Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware reso...Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.展开更多
Traditional meteorological downscaling methods face limitations due to the complex distribution of meteorological variables,which can lead to unstable forecasting results,especially in extreme scenarios.To overcome th...Traditional meteorological downscaling methods face limitations due to the complex distribution of meteorological variables,which can lead to unstable forecasting results,especially in extreme scenarios.To overcome this issue,we propose a convolutional graph neural network(CGNN)model,which we enhance with multilayer feature fusion and a squeeze-and-excitation block.Additionally,we introduce a spatially balanced mean squared error(SBMSE)loss function to address the imbalanced distribution and spatial variability of meteorological variables.The CGNN is capable of extracting essential spatial features and aggregating them from a global perspective,thereby improving the accuracy of prediction and enhancing the model's generalization ability.Based on the experimental results,CGNN has certain advantages in terms of bias distribution,exhibiting a smaller variance.When it comes to precipitation,both UNet and AE also demonstrate relatively small biases.As for temperature,AE and CNNdense perform outstandingly during the winter.The time correlation coefficients show an improvement of at least 10%at daily and monthly scales for both temperature and precipitation.Furthermore,the SBMSE loss function displays an advantage over existing loss functions in predicting the98th percentile and identifying areas where extreme events occur.However,the SBMSE tends to overestimate the distribution of extreme precipitation,which may be due to the theoretical assumptions about the posterior distribution of data that partially limit the effectiveness of the loss function.In future work,we will further optimize the SBMSE to improve prediction accuracy.展开更多
Scene recognition is a popular open problem in the computer vision field.Among lots of methods proposed in recent years,Convolutional Neural Network(CNN)based approaches achieve the best performance in scene recogniti...Scene recognition is a popular open problem in the computer vision field.Among lots of methods proposed in recent years,Convolutional Neural Network(CNN)based approaches achieve the best performance in scene recognition.We propose in this paper an advanced feature fusion algorithm using Multiple Convolutional Neural Network(Multi-CNN)for scene recognition.Unlike existing works that usually use individual convolutional neural network,a fusion of multiple different convolutional neural networks is applied for scene recognition.Firstly,we split training images in two directions and apply to three deep CNN model,and then extract features from the last full-connected(FC)layer and probabilistic layer on each model.Finally,feature vectors are fused with different fusion strategies in groups forwarded into SoftMax classifier.Our proposed algorithm is evaluated on three scene datasets for scene recognition.The experimental results demonstrate the effectiveness of proposed algorithm compared with other state-of-art approaches.展开更多
The complicated electromagnetic environment of the BeiDou satellites introduces vari-ous types of external jamming to communication links,in which recognition of jamming signals with uncertainties is essential.In this...The complicated electromagnetic environment of the BeiDou satellites introduces vari-ous types of external jamming to communication links,in which recognition of jamming signals with uncertainties is essential.In this work,the jamming recognition framework proposed consists of fea-ture fusion and a convolutional neural network(CNN).Firstly,the recognition inputs are obtained by prepossessing procedure,in which the 1-D power spectrum and 2-D time-frequency image are ac-cessed through the Welch algorithm and short-time Fourier transform(STFT),respectively.Then,the 1D-CNN and residual neural network(ResNet)are introduced to extract the deep features of the two prepossessing inputs,respectively.Finally,the two deep features are concatenated for the following three fully connected layers and output the jamming signal classification results through the softmax layer.Results show the proposed method could reduce the impacts of potential feature loss,therefore improving the generalization ability on dealing with uncertainties.展开更多
Ransomware attacks pose a significant threat to critical infrastructures,demanding robust detection mechanisms.This study introduces a hybrid model that combines vision transformer(ViT)and one-dimensional convolutiona...Ransomware attacks pose a significant threat to critical infrastructures,demanding robust detection mechanisms.This study introduces a hybrid model that combines vision transformer(ViT)and one-dimensional convolutional neural network(1DCNN)architectures to enhance ransomware detection capabilities.Addressing common challenges in ransomware detection,particularly dataset class imbalance,the synthetic minority oversampling technique(SMOTE)is employed to generate synthetic samples for minority class,thereby improving detection accuracy.The integration of ViT and 1DCNN through feature fusion enables the model to capture both global contextual and local sequential features,resulting in comprehensive ransomware classification.Tested on the UNSW-NB15 dataset,the proposed ViT-1DCNN model achieved 98%detection accuracy with precision,recall,and F1-score metrics surpassing conventional methods.This approach not only reduces false positives and negatives but also offers scalability and robustness for real-world cybersecurity applications.The results demonstrate the model’s potential as an effective tool for proactive ransomware detection,especially in environments where evolving threats require adaptable and high-accuracy solutions.展开更多
The development of defect prediction plays a significant role in improving software quality. Such predictions are used to identify defective modules before the testing and to minimize the time and cost. The software w...The development of defect prediction plays a significant role in improving software quality. Such predictions are used to identify defective modules before the testing and to minimize the time and cost. The software with defects negatively impacts operational costs and finally affects customer satisfaction. Numerous approaches exist to predict software defects. However, the timely and accurate software bugs are the major challenging issues. To improve the timely and accurate software defect prediction, a novel technique called Nonparametric Statistical feature scaled QuAdratic regressive convolution Deep nEural Network (SQADEN) is introduced. The proposed SQADEN technique mainly includes two major processes namely metric or feature selection and classification. First, the SQADEN uses the nonparametric statistical Torgerson–Gower scaling technique for identifying the relevant software metrics by measuring the similarity using the dice coefficient. The feature selection process is used to minimize the time complexity of software fault prediction. With the selected metrics, software fault perdition with the help of the Quadratic Censored regressive convolution deep neural network-based classification. The deep learning classifier analyzes the training and testing samples using the contingency correlation coefficient. The softstep activation function is used to provide the final fault prediction results. To minimize the error, the Nelder–Mead method is applied to solve non-linear least-squares problems. Finally, accurate classification results with a minimum error are obtained at the output layer. Experimental evaluation is carried out with different quantitative metrics such as accuracy, precision, recall, F-measure, and time complexity. The analyzed results demonstrate the superior performance of our proposed SQADEN technique with maximum accuracy, sensitivity and specificity by 3%, 3%, 2% and 3% and minimum time and space by 13% and 15% when compared with the two state-of-the-art methods.展开更多
In this paper,we utilized the deep convolutional neural network D-LinkNet,a model for semantic segmentation,to analyze the Himawari-8 satellite data captured from 16 channels at a spatial resolution of 0.5 km,with a f...In this paper,we utilized the deep convolutional neural network D-LinkNet,a model for semantic segmentation,to analyze the Himawari-8 satellite data captured from 16 channels at a spatial resolution of 0.5 km,with a focus on the area over the Yellow Sea and the Bohai Sea(32°-42°N,117°-127°E).The objective was to develop an algorithm for fusing and segmenting multi-channel images from geostationary meteorological satellites,specifically for monitoring sea fog in this region.Firstly,the extreme gradient boosting algorithm was adopted to evaluate the data from the 16 channels of the Himawari-8 satellite for sea fog detection,and we found that the top three channels in order of importance were channels 3,4,and 14,which were fused into false color daytime images,while channels 7,13,and 15 were fused into false color nighttime images.Secondly,the simple linear iterative super-pixel clustering algorithm was used for the pixel-level segmentation of false color images,and based on super-pixel blocks,manual sea-fog annotation was performed to obtain fine-grained annotation labels.The deep convolutional neural network D-LinkNet was built on the ResNet backbone and the dilated convolutional layers with direct connections were added in the central part to form a string-and-combine structure with five branches having different depths and receptive fields.Results show that the accuracy rate of fog area(proportion of detected real fog to detected fog)was 66.5%,the recognition rate of fog zone(proportion of detected real fog to real fog or cloud cover)was 51.9%,and the detection accuracy rate(proportion of samples detected correctly to total samples)was 93.2%.展开更多
In recent years, deep convolutional neural networks have shown superior performance in image denoising. However, deep network structures often come with a large number of model parameters, leading to high training cos...In recent years, deep convolutional neural networks have shown superior performance in image denoising. However, deep network structures often come with a large number of model parameters, leading to high training costs and long inference times, limiting their practical application in denoising tasks. This paper proposes a new dual convolutional denoising network with skip connections(DECDNet), which achieves an ideal balance between denoising effect and network complexity. The proposed DECDNet consists of a noise estimation network, a multi-scale feature extraction network, a dual convolutional neural network, and dual attention mechanisms. The noise estimation network is used to estimate the noise level map, and the multi-scale feature extraction network is combined to improve the model's flexibility in obtaining image features. The dual convolutional neural network branch design includes convolution and dilated convolution interactive connections, with the lower branch consisting of dilated convolution layers, and both branches using skip connections. Experiments show that compared with other models, the proposed DECDNet achieves superior PSNR and SSIM values at all compared noise levels, especially at higher noise levels, showing robustness to images with higher noise levels. It also demonstrates better visual effects, maintaining a balance between denoising and detail preservation.展开更多
Spectral and spatial features in remotely sensed data play an irreplaceable role in classifying crop types for precision agriculture. Despite the thriving establishment of the handcrafted features, designing or select...Spectral and spatial features in remotely sensed data play an irreplaceable role in classifying crop types for precision agriculture. Despite the thriving establishment of the handcrafted features, designing or selecting such features valid for specific crop types requires prior knowledge and thus remains an open challenge. Convolutional neural networks(CNNs) can effectively overcome this issue with their advanced ability to generate high-level features automatically but are still inadequate in mining spectral features compared to mining spatial features. This study proposed an enhanced spectral feature called Stacked Spectral Feature Space Patch(SSFSP) for CNN-based crop classification. SSFSP is a stack of twodimensional(2 D) gridded spectral feature images that record various crop types’ spatial and intensity distribution characteristics in a 2 D feature space consisting of two spectral bands. SSFSP can be input into2 D-CNNs to support the simultaneous mining of spectral and spatial features, as the spectral features are successfully converted to 2 D images that can be processed by CNN. We tested the performance of SSFSP by using it as the input to seven CNN models and one multilayer perceptron model for crop type classification compared to using conventional spectral features as input. Using high spatial resolution hyperspectral datasets at three sites, the comparative study demonstrated that SSFSP outperforms conventional spectral features regarding classification accuracy, robustness, and training efficiency. The theoretical analysis summarizes three reasons for its excellent performance. First, SSFSP mines the spectral interrelationship with feature generality, which reduces the required number of training samples.Second, the intra-class variance can be largely reduced by grid partitioning. Third, SSFSP is a highly sparse feature, which reduces the dependence on the CNN model structure and enables early and fast convergence in model training. In conclusion, SSFSP has great potential for practical crop classification in precision agriculture.展开更多
To improve the quality of the infrared image and enhance the information of the object,a dual band infrared image fusion method based on feature extraction and a novel multiple pulse coupled neural network(multi-PCNN)...To improve the quality of the infrared image and enhance the information of the object,a dual band infrared image fusion method based on feature extraction and a novel multiple pulse coupled neural network(multi-PCNN)is proposed.In this multi-PCNN fusion scheme,the auxiliary PCNN which captures the characteristics of feature image extracting from the infrared image is used to modulate the main PCNN,whose input could be original infrared image.Meanwhile,to make the PCNN fusion effect consistent with the human vision system,Laplacian energy is adopted to obtain the value of adaptive linking strength in PCNN.After that,the original dual band infrared images are reconstructed by using a weight fusion rule with the fire mapping images generated by the main PCNNs to obtain the fused image.Compared to wavelet transforms,Laplacian pyramids and traditional multi-PCNNs,fusion images based on our method have more information,rich details and clear edges.展开更多
Faced with the massive amount of online shopping clothing images,how to classify them quickly and accurately is a challenging task in image classification.In this paper,we propose a novel method,named Multi_XMNet,to s...Faced with the massive amount of online shopping clothing images,how to classify them quickly and accurately is a challenging task in image classification.In this paper,we propose a novel method,named Multi_XMNet,to solve the clothing images classification problem.The proposed method mainly consists of two convolution neural network(CNN)branches.One branch extracts multiscale features from the whole expressional image by Multi_X which is designed by improving the Xception network,while the other extracts attention mechanism features from the whole expressional image by MobileNetV3-small network.Both multiscale and attention mechanism features are aggregated before making classification.Additionally,in the training stage,global average pooling(GAP),convolutional layers,and softmax classifiers are used instead of the fully connected layer to classify the final features,which speed up model training and alleviate the problem of overfitting caused by too many parameters.Experimental comparisons are made in the public DeepFashion dataset.The experimental results show that the classification accuracy of this method is 95.38%,which is better than InceptionV3,Xception and InceptionV3_Xception by 5.58%,3.32%,and 2.22%,respectively.The proposed Multi_XMNet image classification model can help enterprises and researchers in the field of clothing e-commerce to automaticly,efficiently and accurately classify massive clothing images.展开更多
<div style="text-align:justify;"> Load identification method is one of the major technical difficulties of non-intrusive composite monitoring. Binary V-I trajectory image can reflect the original V-I t...<div style="text-align:justify;"> Load identification method is one of the major technical difficulties of non-intrusive composite monitoring. Binary V-I trajectory image can reflect the original V-I trajectory characteristics to a large extent, so it is widely used in load identification. However, using single binary V-I trajectory feature for load identification has certain limitations. In order to improve the accuracy of load identification, the power feature is added on the basis of the binary V-I trajectory feature in this paper. We change the initial binary V-I trajectory into a new 3D feature by mapping the power feature to the third dimension. In order to reduce the impact of imbalance samples on load identification, the SVM SMOTE algorithm is used to balance the samples. Based on the deep learning method, the convolutional neural network model is used to extract the newly produced 3D feature to achieve load identification in this paper. The results indicate the new 3D feature has better observability and the proposed model has higher identification performance compared with other classification models on the public data set PLAID. </div>展开更多
Recent advancements have established machine learning's utility in predicting nonlinear fluid dynamics,with predictive accuracy being a central motivation for employing neural networks.However,the pattern recognit...Recent advancements have established machine learning's utility in predicting nonlinear fluid dynamics,with predictive accuracy being a central motivation for employing neural networks.However,the pattern recognition central to the networks function is equally valuable for enhancing our dynamical insight into the complex fluid dynamics.In this paper,a single-layer convolutional neural network(CNN)was trained to recognize three qualitatively different subsonic buffet flows(periodic,quasi-periodic and chaotic)over a high-incidence airfoil,and a near-perfect accuracy was obtained with only a small training dataset.The convolutional kernels and corresponding feature maps,developed by the model with no temporal information provided,identified large-scale coherent structures in agreement with those known to be associated with buffet flows.Sensitivity to hyperparameters including network architecture and convolutional kernel size was also explored.The coherent structures identified by these models enhance our dynamical understanding of subsonic buffet over high-incidence airfoils over a wide range of Reynolds numbers.展开更多
With the continuous progress of The Times and the development of technology,the rise of network social media has also brought the“explosive”growth of image data.As one of the main ways of People’s Daily communicati...With the continuous progress of The Times and the development of technology,the rise of network social media has also brought the“explosive”growth of image data.As one of the main ways of People’s Daily communication,image is widely used as a carrier of communication because of its rich content,intuitive and other advantages.Image recognition based on convolution neural network is the first application in the field of image recognition.A series of algorithm operations such as image eigenvalue extraction,recognition and convolution are used to identify and analyze different images.The rapid development of artificial intelligence makes machine learning more and more important in its research field.Use algorithms to learn each piece of data and predict the outcome.This has become an important key to open the door of artificial intelligence.In machine vision,image recognition is the foundation,but how to associate the low-level information in the image with the high-level image semantics becomes the key problem of image recognition.Predecessors have provided many model algorithms,which have laid a solid foundation for the development of artificial intelligence and image recognition.The multi-level information fusion model based on the VGG16 model is an improvement on the fully connected neural network.Different from full connection network,convolutional neural network does not use full connection method in each layer of neurons of neural network,but USES some nodes for connection.Although this method reduces the computation time,due to the fact that the convolutional neural network model will lose some useful feature information in the process of propagation and calculation,this paper improves the model to be a multi-level information fusion of the convolution calculation method,and further recovers the discarded feature information,so as to improve the recognition rate of the image.VGG divides the network into five groups(mimicking the five layers of AlexNet),yet it USES 3*3 filters and combines them as a convolution sequence.Network deeper DCNN,channel number is bigger.The recognition rate of the model was verified by 0RL Face Database,BioID Face Database and CASIA Face Image Database.展开更多
With the development of deep learning and Convolutional Neural Networks(CNNs),the accuracy of automatic food recognition based on visual data have significantly improved.Some research studies have shown that the deepe...With the development of deep learning and Convolutional Neural Networks(CNNs),the accuracy of automatic food recognition based on visual data have significantly improved.Some research studies have shown that the deeper the model is,the higher the accuracy is.However,very deep neural networks would be affected by the overfitting problem and also consume huge computing resources.In this paper,a new classification scheme is proposed for automatic food-ingredient recognition based on deep learning.We construct an up-to-date combinational convolutional neural network(CBNet)with a subnet merging technique.Firstly,two different neural networks are utilized for learning interested features.Then,a well-designed feature fusion component aggregates the features from subnetworks,further extracting richer and more precise features for image classification.In order to learn more complementary features,the corresponding fusion strategies are also proposed,including auxiliary classifiers and hyperparameters setting.Finally,CBNet based on the well-known VGGNet,ResNet and DenseNet is evaluated on a dataset including 41 major categories of food ingredients and 100 images for each category.Theoretical analysis and experimental results demonstrate that CBNet achieves promising accuracy for multi-class classification and improves the performance of convolutional neural networks.展开更多
Image retrieval for food ingredients is important work,tremendously tiring,uninteresting,and expensive.Computer vision systems have extraordinary advancements in image retrieval with CNNs skills.But it is not feasible...Image retrieval for food ingredients is important work,tremendously tiring,uninteresting,and expensive.Computer vision systems have extraordinary advancements in image retrieval with CNNs skills.But it is not feasible for small-size food datasets using convolutional neural networks directly.In this study,a novel image retrieval approach is presented for small and medium-scale food datasets,which both augments images utilizing image transformation techniques to enlarge the size of datasets,and promotes the average accuracy of food recognition with state-of-the-art deep learning technologies.First,typical image transformation techniques are used to augment food images.Then transfer learning technology based on deep learning is applied to extract image features.Finally,a food recognition algorithm is leveraged on extracted deepfeature vectors.The presented image-retrieval architecture is analyzed based on a smallscale food dataset which is composed of forty-one categories of food ingredients and one hundred pictures for each category.Extensive experimental results demonstrate the advantages of image-augmentation architecture for small and medium datasets using deep learning.The novel approach combines image augmentation,ResNet feature vectors,and SMO classification,and shows its superiority for food detection of small/medium-scale datasets with comprehensive experiments.展开更多
The widespread availability of digital multimedia data has led to a new challenge in digital forensics.Traditional source camera identification algorithms usually rely on various traces in the capturing process.Howeve...The widespread availability of digital multimedia data has led to a new challenge in digital forensics.Traditional source camera identification algorithms usually rely on various traces in the capturing process.However,these traces have become increasingly difficult to extract due to wide availability of various image processing algorithms.Convolutional Neural Networks(CNN)-based algorithms have demonstrated good discriminative capabilities for different brands and even different models of camera devices.However,their performances is not ideal in case of distinguishing between individual devices of the same model,because cameras of the same model typically use the same optical lens,image sensor,and image processing algorithms,that result in minimal overall differences.In this paper,we propose a camera forensics algorithm based on multi-scale feature fusion to address these issues.The proposed algorithm extracts different local features from feature maps of different scales and then fuses them to obtain a comprehensive feature representation.This representation is then fed into a subsequent camera fingerprint classification network.Building upon the Swin-T network,we utilize Transformer Blocks and Graph Convolutional Network(GCN)modules to fuse multi-scale features from different stages of the backbone network.Furthermore,we conduct experiments on established datasets to demonstrate the feasibility and effectiveness of the proposed approach.展开更多
With the rapid development of the mobile communication and the Internet,the previous web anomaly detectionand identificationmodels were built relying on security experts’empirical knowledge and attack features.Althou...With the rapid development of the mobile communication and the Internet,the previous web anomaly detectionand identificationmodels were built relying on security experts’empirical knowledge and attack features.Althoughthis approach can achieve higher detection performance,it requires huge human labor and resources to maintainthe feature library.In contrast,semantic feature engineering can dynamically discover new semantic featuresand optimize feature selection by automatically analyzing the semantic information contained in the data itself,thus reducing dependence on prior knowledge.However,current semantic features still have the problem ofsemantic expression singularity,as they are extracted from a single semantic mode such as word segmentation,character segmentation,or arbitrary semantic feature extraction.This paper extracts features of web requestsfrom dual semantic granularity,and proposes a semantic feature fusion method to solve the above problems.Themethod first preprocesses web requests,and extracts word-level and character-level semantic features of URLs viaconvolutional neural network(CNN),respectively.By constructing three loss functions to reduce losses betweenfeatures,labels and categories.Experiments on the HTTP CSIC 2010,Malicious URLs and HttpParams datasetsverify the proposedmethod.Results show that compared withmachine learning,deep learningmethods and BERTmodel,the proposed method has better detection performance.And it achieved the best detection rate of 99.16%in the dataset HttpParams.展开更多
Automatic segmentation of medical images provides a reliable scientific basis for disease diagnosis and analysis.Notably,most existing methods that combine the strengths of convolutional neural networks(CNNs)and Trans...Automatic segmentation of medical images provides a reliable scientific basis for disease diagnosis and analysis.Notably,most existing methods that combine the strengths of convolutional neural networks(CNNs)and Transformers have made significant progress.However,there are some limitations in the current integration of CNN and Transformer technology in two key aspects.Firstly,most methods either overlook or fail to fully incorporate the complementary nature between local and global features.Secondly,the significance of integrating the multiscale encoder features from the dual-branch network to enhance the decoding features is often disregarded in methods that combine CNN and Transformer.To address this issue,we present a groundbreaking dual-branch cross-attention fusion network(DCFNet),which efficiently combines the power of Swin Transformer and CNN to generate complementary global and local features.We then designed the Feature Cross-Fusion(FCF)module to efficiently fuse local and global features.In the FCF,the utilization of the Channel-wise Cross-fusion Transformer(CCT)serves the purpose of aggregatingmulti-scale features,and the Feature FusionModule(FFM)is employed to effectively aggregate dual-branch prominent feature regions from the spatial perspective.Furthermore,within the decoding phase of the dual-branch network,our proposed Channel Attention Block(CAB)aims to emphasize the significance of the channel features between the up-sampled features and the features generated by the FCFmodule to enhance the details of the decoding.Experimental results demonstrate that DCFNet exhibits enhanced accuracy in segmentation performance.Compared to other state-of-the-art(SOTA)methods,our segmentation framework exhibits a superior level of competitiveness.DCFNet’s accurate segmentation of medical images can greatly assist medical professionals in making crucial diagnoses of lesion areas in advance.展开更多
In order to improve the detection accuracy of small objects,a neighborhood fusion-based hierarchical parallel feature pyramid network(NFPN)is proposed.Unlike the layer-by-layer structure adopted in the feature pyramid...In order to improve the detection accuracy of small objects,a neighborhood fusion-based hierarchical parallel feature pyramid network(NFPN)is proposed.Unlike the layer-by-layer structure adopted in the feature pyramid network(FPN)and deconvolutional single shot detector(DSSD),where the bottom layer of the feature pyramid network relies on the top layer,NFPN builds the feature pyramid network with no connections between the upper and lower layers.That is,it only fuses shallow features on similar scales.NFPN is highly portable and can be embedded in many models to further boost performance.Extensive experiments on PASCAL VOC 2007,2012,and COCO datasets demonstrate that the NFPN-based SSD without intricate tricks can exceed the DSSD model in terms of detection accuracy and inference speed,especially for small objects,e.g.,4%to 5%higher mAP(mean average precision)than SSD,and 2%to 3%higher mAP than DSSD.On VOC 2007 test set,the NFPN-based SSD with 300×300 input reaches 79.4%mAP at 34.6 frame/s,and the mAP can raise to 82.9%after using the multi-scale testing strategy.展开更多
文摘Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.
基金partially funded by the National Natural Science Foundation of China(U2142205)the Guangdong Major Project of Basic and Applied Basic Research(2020B0301030004)+1 种基金the Special Fund for Forecasters of China Meteorological Administration(CMAYBY2020-094)the Graduate Student Research and Innovation Program of Central South University(2023ZZTS0347)。
文摘Traditional meteorological downscaling methods face limitations due to the complex distribution of meteorological variables,which can lead to unstable forecasting results,especially in extreme scenarios.To overcome this issue,we propose a convolutional graph neural network(CGNN)model,which we enhance with multilayer feature fusion and a squeeze-and-excitation block.Additionally,we introduce a spatially balanced mean squared error(SBMSE)loss function to address the imbalanced distribution and spatial variability of meteorological variables.The CGNN is capable of extracting essential spatial features and aggregating them from a global perspective,thereby improving the accuracy of prediction and enhancing the model's generalization ability.Based on the experimental results,CGNN has certain advantages in terms of bias distribution,exhibiting a smaller variance.When it comes to precipitation,both UNet and AE also demonstrate relatively small biases.As for temperature,AE and CNNdense perform outstandingly during the winter.The time correlation coefficients show an improvement of at least 10%at daily and monthly scales for both temperature and precipitation.Furthermore,the SBMSE loss function displays an advantage over existing loss functions in predicting the98th percentile and identifying areas where extreme events occur.However,the SBMSE tends to overestimate the distribution of extreme precipitation,which may be due to the theoretical assumptions about the posterior distribution of data that partially limit the effectiveness of the loss function.In future work,we will further optimize the SBMSE to improve prediction accuracy.
文摘Scene recognition is a popular open problem in the computer vision field.Among lots of methods proposed in recent years,Convolutional Neural Network(CNN)based approaches achieve the best performance in scene recognition.We propose in this paper an advanced feature fusion algorithm using Multiple Convolutional Neural Network(Multi-CNN)for scene recognition.Unlike existing works that usually use individual convolutional neural network,a fusion of multiple different convolutional neural networks is applied for scene recognition.Firstly,we split training images in two directions and apply to three deep CNN model,and then extract features from the last full-connected(FC)layer and probabilistic layer on each model.Finally,feature vectors are fused with different fusion strategies in groups forwarded into SoftMax classifier.Our proposed algorithm is evaluated on three scene datasets for scene recognition.The experimental results demonstrate the effectiveness of proposed algorithm compared with other state-of-art approaches.
基金supported by the National Key Research and De-velopment Program of China(No.2020YFB0505601).
文摘The complicated electromagnetic environment of the BeiDou satellites introduces vari-ous types of external jamming to communication links,in which recognition of jamming signals with uncertainties is essential.In this work,the jamming recognition framework proposed consists of fea-ture fusion and a convolutional neural network(CNN).Firstly,the recognition inputs are obtained by prepossessing procedure,in which the 1-D power spectrum and 2-D time-frequency image are ac-cessed through the Welch algorithm and short-time Fourier transform(STFT),respectively.Then,the 1D-CNN and residual neural network(ResNet)are introduced to extract the deep features of the two prepossessing inputs,respectively.Finally,the two deep features are concatenated for the following three fully connected layers and output the jamming signal classification results through the softmax layer.Results show the proposed method could reduce the impacts of potential feature loss,therefore improving the generalization ability on dealing with uncertainties.
文摘Ransomware attacks pose a significant threat to critical infrastructures,demanding robust detection mechanisms.This study introduces a hybrid model that combines vision transformer(ViT)and one-dimensional convolutional neural network(1DCNN)architectures to enhance ransomware detection capabilities.Addressing common challenges in ransomware detection,particularly dataset class imbalance,the synthetic minority oversampling technique(SMOTE)is employed to generate synthetic samples for minority class,thereby improving detection accuracy.The integration of ViT and 1DCNN through feature fusion enables the model to capture both global contextual and local sequential features,resulting in comprehensive ransomware classification.Tested on the UNSW-NB15 dataset,the proposed ViT-1DCNN model achieved 98%detection accuracy with precision,recall,and F1-score metrics surpassing conventional methods.This approach not only reduces false positives and negatives but also offers scalability and robustness for real-world cybersecurity applications.The results demonstrate the model’s potential as an effective tool for proactive ransomware detection,especially in environments where evolving threats require adaptable and high-accuracy solutions.
文摘The development of defect prediction plays a significant role in improving software quality. Such predictions are used to identify defective modules before the testing and to minimize the time and cost. The software with defects negatively impacts operational costs and finally affects customer satisfaction. Numerous approaches exist to predict software defects. However, the timely and accurate software bugs are the major challenging issues. To improve the timely and accurate software defect prediction, a novel technique called Nonparametric Statistical feature scaled QuAdratic regressive convolution Deep nEural Network (SQADEN) is introduced. The proposed SQADEN technique mainly includes two major processes namely metric or feature selection and classification. First, the SQADEN uses the nonparametric statistical Torgerson–Gower scaling technique for identifying the relevant software metrics by measuring the similarity using the dice coefficient. The feature selection process is used to minimize the time complexity of software fault prediction. With the selected metrics, software fault perdition with the help of the Quadratic Censored regressive convolution deep neural network-based classification. The deep learning classifier analyzes the training and testing samples using the contingency correlation coefficient. The softstep activation function is used to provide the final fault prediction results. To minimize the error, the Nelder–Mead method is applied to solve non-linear least-squares problems. Finally, accurate classification results with a minimum error are obtained at the output layer. Experimental evaluation is carried out with different quantitative metrics such as accuracy, precision, recall, F-measure, and time complexity. The analyzed results demonstrate the superior performance of our proposed SQADEN technique with maximum accuracy, sensitivity and specificity by 3%, 3%, 2% and 3% and minimum time and space by 13% and 15% when compared with the two state-of-the-art methods.
基金National Key R&D Program of China(2021YFC3000905)Open Research Program of the State Key Laboratory of Severe Weather(2022LASW-B09)National Natural Science Foundation of China(42375010)。
文摘In this paper,we utilized the deep convolutional neural network D-LinkNet,a model for semantic segmentation,to analyze the Himawari-8 satellite data captured from 16 channels at a spatial resolution of 0.5 km,with a focus on the area over the Yellow Sea and the Bohai Sea(32°-42°N,117°-127°E).The objective was to develop an algorithm for fusing and segmenting multi-channel images from geostationary meteorological satellites,specifically for monitoring sea fog in this region.Firstly,the extreme gradient boosting algorithm was adopted to evaluate the data from the 16 channels of the Himawari-8 satellite for sea fog detection,and we found that the top three channels in order of importance were channels 3,4,and 14,which were fused into false color daytime images,while channels 7,13,and 15 were fused into false color nighttime images.Secondly,the simple linear iterative super-pixel clustering algorithm was used for the pixel-level segmentation of false color images,and based on super-pixel blocks,manual sea-fog annotation was performed to obtain fine-grained annotation labels.The deep convolutional neural network D-LinkNet was built on the ResNet backbone and the dilated convolutional layers with direct connections were added in the central part to form a string-and-combine structure with five branches having different depths and receptive fields.Results show that the accuracy rate of fog area(proportion of detected real fog to detected fog)was 66.5%,the recognition rate of fog zone(proportion of detected real fog to real fog or cloud cover)was 51.9%,and the detection accuracy rate(proportion of samples detected correctly to total samples)was 93.2%.
基金funded by National Nature Science Foundation of China,grant number 61302188。
文摘In recent years, deep convolutional neural networks have shown superior performance in image denoising. However, deep network structures often come with a large number of model parameters, leading to high training costs and long inference times, limiting their practical application in denoising tasks. This paper proposes a new dual convolutional denoising network with skip connections(DECDNet), which achieves an ideal balance between denoising effect and network complexity. The proposed DECDNet consists of a noise estimation network, a multi-scale feature extraction network, a dual convolutional neural network, and dual attention mechanisms. The noise estimation network is used to estimate the noise level map, and the multi-scale feature extraction network is combined to improve the model's flexibility in obtaining image features. The dual convolutional neural network branch design includes convolution and dilated convolution interactive connections, with the lower branch consisting of dilated convolution layers, and both branches using skip connections. Experiments show that compared with other models, the proposed DECDNet achieves superior PSNR and SSIM values at all compared noise levels, especially at higher noise levels, showing robustness to images with higher noise levels. It also demonstrates better visual effects, maintaining a balance between denoising and detail preservation.
基金supported by the National Natural Science Foundation of China (67441830108 and 41871224)。
文摘Spectral and spatial features in remotely sensed data play an irreplaceable role in classifying crop types for precision agriculture. Despite the thriving establishment of the handcrafted features, designing or selecting such features valid for specific crop types requires prior knowledge and thus remains an open challenge. Convolutional neural networks(CNNs) can effectively overcome this issue with their advanced ability to generate high-level features automatically but are still inadequate in mining spectral features compared to mining spatial features. This study proposed an enhanced spectral feature called Stacked Spectral Feature Space Patch(SSFSP) for CNN-based crop classification. SSFSP is a stack of twodimensional(2 D) gridded spectral feature images that record various crop types’ spatial and intensity distribution characteristics in a 2 D feature space consisting of two spectral bands. SSFSP can be input into2 D-CNNs to support the simultaneous mining of spectral and spatial features, as the spectral features are successfully converted to 2 D images that can be processed by CNN. We tested the performance of SSFSP by using it as the input to seven CNN models and one multilayer perceptron model for crop type classification compared to using conventional spectral features as input. Using high spatial resolution hyperspectral datasets at three sites, the comparative study demonstrated that SSFSP outperforms conventional spectral features regarding classification accuracy, robustness, and training efficiency. The theoretical analysis summarizes three reasons for its excellent performance. First, SSFSP mines the spectral interrelationship with feature generality, which reduces the required number of training samples.Second, the intra-class variance can be largely reduced by grid partitioning. Third, SSFSP is a highly sparse feature, which reduces the dependence on the CNN model structure and enables early and fast convergence in model training. In conclusion, SSFSP has great potential for practical crop classification in precision agriculture.
基金Supported by the National Natural Science Foundation of China(60905012,60572058)
文摘To improve the quality of the infrared image and enhance the information of the object,a dual band infrared image fusion method based on feature extraction and a novel multiple pulse coupled neural network(multi-PCNN)is proposed.In this multi-PCNN fusion scheme,the auxiliary PCNN which captures the characteristics of feature image extracting from the infrared image is used to modulate the main PCNN,whose input could be original infrared image.Meanwhile,to make the PCNN fusion effect consistent with the human vision system,Laplacian energy is adopted to obtain the value of adaptive linking strength in PCNN.After that,the original dual band infrared images are reconstructed by using a weight fusion rule with the fire mapping images generated by the main PCNNs to obtain the fused image.Compared to wavelet transforms,Laplacian pyramids and traditional multi-PCNNs,fusion images based on our method have more information,rich details and clear edges.
基金Fundamental Research Funds for the Central Universities of Ministry of Education of China(No.19D111201)。
文摘Faced with the massive amount of online shopping clothing images,how to classify them quickly and accurately is a challenging task in image classification.In this paper,we propose a novel method,named Multi_XMNet,to solve the clothing images classification problem.The proposed method mainly consists of two convolution neural network(CNN)branches.One branch extracts multiscale features from the whole expressional image by Multi_X which is designed by improving the Xception network,while the other extracts attention mechanism features from the whole expressional image by MobileNetV3-small network.Both multiscale and attention mechanism features are aggregated before making classification.Additionally,in the training stage,global average pooling(GAP),convolutional layers,and softmax classifiers are used instead of the fully connected layer to classify the final features,which speed up model training and alleviate the problem of overfitting caused by too many parameters.Experimental comparisons are made in the public DeepFashion dataset.The experimental results show that the classification accuracy of this method is 95.38%,which is better than InceptionV3,Xception and InceptionV3_Xception by 5.58%,3.32%,and 2.22%,respectively.The proposed Multi_XMNet image classification model can help enterprises and researchers in the field of clothing e-commerce to automaticly,efficiently and accurately classify massive clothing images.
文摘<div style="text-align:justify;"> Load identification method is one of the major technical difficulties of non-intrusive composite monitoring. Binary V-I trajectory image can reflect the original V-I trajectory characteristics to a large extent, so it is widely used in load identification. However, using single binary V-I trajectory feature for load identification has certain limitations. In order to improve the accuracy of load identification, the power feature is added on the basis of the binary V-I trajectory feature in this paper. We change the initial binary V-I trajectory into a new 3D feature by mapping the power feature to the third dimension. In order to reduce the impact of imbalance samples on load identification, the SVM SMOTE algorithm is used to balance the samples. Based on the deep learning method, the convolutional neural network model is used to extract the newly produced 3D feature to achieve load identification in this paper. The results indicate the new 3D feature has better observability and the proposed model has higher identification performance compared with other classification models on the public data set PLAID. </div>
文摘Recent advancements have established machine learning's utility in predicting nonlinear fluid dynamics,with predictive accuracy being a central motivation for employing neural networks.However,the pattern recognition central to the networks function is equally valuable for enhancing our dynamical insight into the complex fluid dynamics.In this paper,a single-layer convolutional neural network(CNN)was trained to recognize three qualitatively different subsonic buffet flows(periodic,quasi-periodic and chaotic)over a high-incidence airfoil,and a near-perfect accuracy was obtained with only a small training dataset.The convolutional kernels and corresponding feature maps,developed by the model with no temporal information provided,identified large-scale coherent structures in agreement with those known to be associated with buffet flows.Sensitivity to hyperparameters including network architecture and convolutional kernel size was also explored.The coherent structures identified by these models enhance our dynamical understanding of subsonic buffet over high-incidence airfoils over a wide range of Reynolds numbers.
文摘With the continuous progress of The Times and the development of technology,the rise of network social media has also brought the“explosive”growth of image data.As one of the main ways of People’s Daily communication,image is widely used as a carrier of communication because of its rich content,intuitive and other advantages.Image recognition based on convolution neural network is the first application in the field of image recognition.A series of algorithm operations such as image eigenvalue extraction,recognition and convolution are used to identify and analyze different images.The rapid development of artificial intelligence makes machine learning more and more important in its research field.Use algorithms to learn each piece of data and predict the outcome.This has become an important key to open the door of artificial intelligence.In machine vision,image recognition is the foundation,but how to associate the low-level information in the image with the high-level image semantics becomes the key problem of image recognition.Predecessors have provided many model algorithms,which have laid a solid foundation for the development of artificial intelligence and image recognition.The multi-level information fusion model based on the VGG16 model is an improvement on the fully connected neural network.Different from full connection network,convolutional neural network does not use full connection method in each layer of neurons of neural network,but USES some nodes for connection.Although this method reduces the computation time,due to the fact that the convolutional neural network model will lose some useful feature information in the process of propagation and calculation,this paper improves the model to be a multi-level information fusion of the convolution calculation method,and further recovers the discarded feature information,so as to improve the recognition rate of the image.VGG divides the network into five groups(mimicking the five layers of AlexNet),yet it USES 3*3 filters and combines them as a convolution sequence.Network deeper DCNN,channel number is bigger.The recognition rate of the model was verified by 0RL Face Database,BioID Face Database and CASIA Face Image Database.
基金This paper is partially supported by National Natural Foundation of China(Grant No.61772561)the Key Research&Development Plan of Hunan Province(Grant No.2018NK2012)+2 种基金Postgraduate Research and Innovative Project of Central South University of Forestry and Technology(Grant No.20183012)Graduate Education and Teaching Reform Project of Central South University of Forestry and Technology(Grant No.2018JG005)Teaching Reform Project of Central South University of Forestry and Technology(Grant No.20180682).
文摘With the development of deep learning and Convolutional Neural Networks(CNNs),the accuracy of automatic food recognition based on visual data have significantly improved.Some research studies have shown that the deeper the model is,the higher the accuracy is.However,very deep neural networks would be affected by the overfitting problem and also consume huge computing resources.In this paper,a new classification scheme is proposed for automatic food-ingredient recognition based on deep learning.We construct an up-to-date combinational convolutional neural network(CBNet)with a subnet merging technique.Firstly,two different neural networks are utilized for learning interested features.Then,a well-designed feature fusion component aggregates the features from subnetworks,further extracting richer and more precise features for image classification.In order to learn more complementary features,the corresponding fusion strategies are also proposed,including auxiliary classifiers and hyperparameters setting.Finally,CBNet based on the well-known VGGNet,ResNet and DenseNet is evaluated on a dataset including 41 major categories of food ingredients and 100 images for each category.Theoretical analysis and experimental results demonstrate that CBNet achieves promising accuracy for multi-class classification and improves the performance of convolutional neural networks.
文摘Image retrieval for food ingredients is important work,tremendously tiring,uninteresting,and expensive.Computer vision systems have extraordinary advancements in image retrieval with CNNs skills.But it is not feasible for small-size food datasets using convolutional neural networks directly.In this study,a novel image retrieval approach is presented for small and medium-scale food datasets,which both augments images utilizing image transformation techniques to enlarge the size of datasets,and promotes the average accuracy of food recognition with state-of-the-art deep learning technologies.First,typical image transformation techniques are used to augment food images.Then transfer learning technology based on deep learning is applied to extract image features.Finally,a food recognition algorithm is leveraged on extracted deepfeature vectors.The presented image-retrieval architecture is analyzed based on a smallscale food dataset which is composed of forty-one categories of food ingredients and one hundred pictures for each category.Extensive experimental results demonstrate the advantages of image-augmentation architecture for small and medium datasets using deep learning.The novel approach combines image augmentation,ResNet feature vectors,and SMO classification,and shows its superiority for food detection of small/medium-scale datasets with comprehensive experiments.
基金This work was funded by the National Natural Science Foundation of China(Grant No.62172132)Public Welfare Technology Research Project of Zhejiang Province(Grant No.LGF21F020014)the Opening Project of Key Laboratory of Public Security Information Application Based on Big-Data Architecture,Ministry of Public Security of Zhejiang Police College(Grant No.2021DSJSYS002).
文摘The widespread availability of digital multimedia data has led to a new challenge in digital forensics.Traditional source camera identification algorithms usually rely on various traces in the capturing process.However,these traces have become increasingly difficult to extract due to wide availability of various image processing algorithms.Convolutional Neural Networks(CNN)-based algorithms have demonstrated good discriminative capabilities for different brands and even different models of camera devices.However,their performances is not ideal in case of distinguishing between individual devices of the same model,because cameras of the same model typically use the same optical lens,image sensor,and image processing algorithms,that result in minimal overall differences.In this paper,we propose a camera forensics algorithm based on multi-scale feature fusion to address these issues.The proposed algorithm extracts different local features from feature maps of different scales and then fuses them to obtain a comprehensive feature representation.This representation is then fed into a subsequent camera fingerprint classification network.Building upon the Swin-T network,we utilize Transformer Blocks and Graph Convolutional Network(GCN)modules to fuse multi-scale features from different stages of the backbone network.Furthermore,we conduct experiments on established datasets to demonstrate the feasibility and effectiveness of the proposed approach.
基金a grant from the National Natural Science Foundation of China(Nos.11905239,12005248 and 12105303).
文摘With the rapid development of the mobile communication and the Internet,the previous web anomaly detectionand identificationmodels were built relying on security experts’empirical knowledge and attack features.Althoughthis approach can achieve higher detection performance,it requires huge human labor and resources to maintainthe feature library.In contrast,semantic feature engineering can dynamically discover new semantic featuresand optimize feature selection by automatically analyzing the semantic information contained in the data itself,thus reducing dependence on prior knowledge.However,current semantic features still have the problem ofsemantic expression singularity,as they are extracted from a single semantic mode such as word segmentation,character segmentation,or arbitrary semantic feature extraction.This paper extracts features of web requestsfrom dual semantic granularity,and proposes a semantic feature fusion method to solve the above problems.Themethod first preprocesses web requests,and extracts word-level and character-level semantic features of URLs viaconvolutional neural network(CNN),respectively.By constructing three loss functions to reduce losses betweenfeatures,labels and categories.Experiments on the HTTP CSIC 2010,Malicious URLs and HttpParams datasetsverify the proposedmethod.Results show that compared withmachine learning,deep learningmethods and BERTmodel,the proposed method has better detection performance.And it achieved the best detection rate of 99.16%in the dataset HttpParams.
基金supported by the National Key R&D Program of China(2018AAA0102100)the National Natural Science Foundation of China(No.62376287)+3 种基金the International Science and Technology Innovation Joint Base of Machine Vision and Medical Image Processing in Hunan Province(2021CB1013)the Key Research and Development Program of Hunan Province(2022SK2054)the Natural Science Foundation of Hunan Province(No.2022JJ30762,2023JJ70016)the 111 Project under Grant(No.B18059).
文摘Automatic segmentation of medical images provides a reliable scientific basis for disease diagnosis and analysis.Notably,most existing methods that combine the strengths of convolutional neural networks(CNNs)and Transformers have made significant progress.However,there are some limitations in the current integration of CNN and Transformer technology in two key aspects.Firstly,most methods either overlook or fail to fully incorporate the complementary nature between local and global features.Secondly,the significance of integrating the multiscale encoder features from the dual-branch network to enhance the decoding features is often disregarded in methods that combine CNN and Transformer.To address this issue,we present a groundbreaking dual-branch cross-attention fusion network(DCFNet),which efficiently combines the power of Swin Transformer and CNN to generate complementary global and local features.We then designed the Feature Cross-Fusion(FCF)module to efficiently fuse local and global features.In the FCF,the utilization of the Channel-wise Cross-fusion Transformer(CCT)serves the purpose of aggregatingmulti-scale features,and the Feature FusionModule(FFM)is employed to effectively aggregate dual-branch prominent feature regions from the spatial perspective.Furthermore,within the decoding phase of the dual-branch network,our proposed Channel Attention Block(CAB)aims to emphasize the significance of the channel features between the up-sampled features and the features generated by the FCFmodule to enhance the details of the decoding.Experimental results demonstrate that DCFNet exhibits enhanced accuracy in segmentation performance.Compared to other state-of-the-art(SOTA)methods,our segmentation framework exhibits a superior level of competitiveness.DCFNet’s accurate segmentation of medical images can greatly assist medical professionals in making crucial diagnoses of lesion areas in advance.
基金The National Natural Science Foundation of China(No.61603091)。
文摘In order to improve the detection accuracy of small objects,a neighborhood fusion-based hierarchical parallel feature pyramid network(NFPN)is proposed.Unlike the layer-by-layer structure adopted in the feature pyramid network(FPN)and deconvolutional single shot detector(DSSD),where the bottom layer of the feature pyramid network relies on the top layer,NFPN builds the feature pyramid network with no connections between the upper and lower layers.That is,it only fuses shallow features on similar scales.NFPN is highly portable and can be embedded in many models to further boost performance.Extensive experiments on PASCAL VOC 2007,2012,and COCO datasets demonstrate that the NFPN-based SSD without intricate tricks can exceed the DSSD model in terms of detection accuracy and inference speed,especially for small objects,e.g.,4%to 5%higher mAP(mean average precision)than SSD,and 2%to 3%higher mAP than DSSD.On VOC 2007 test set,the NFPN-based SSD with 300×300 input reaches 79.4%mAP at 34.6 frame/s,and the mAP can raise to 82.9%after using the multi-scale testing strategy.
