摘要
为了应对矿产领域的矿石识别困难、识别设备成本高等问题提出了改进的Unet矿石图像分割算法,首先对EfficientNetV2加以改造作为模型的主干网络提取矿石特征;其次,引入MBconv模块作为解码器中的基本模块,加强特征提取能力;采用CA注意力模块替换SE注意力模块,保留更多空间位置信息;最后使用滤波器响应归一化(FRN)层替换常用的批量归一化(BN)层,避免模型性能受批量大小的影响。基于FeM和Cu数据集实验结果表明,PA分别为96.58%和95.39%,Miou分别为92.8%和90.43%,F1_score分别为95.15%和93.47%,Efficient_unet模型参数相比Unet减少了60.4%,推理速度提升了19.23%,可达21.7 fps,所提模型在精度和速率方面均优于现有经典分割模型,同时表现出良好的泛化性能。
Addressing the challenges of ore recognition in the mining industry and the high cost of recognition equipment,we propose an improved Unet ore image segmentation algorithm.Firstly,we modify EfficientNetV2as the backbone network of the model to extract ore features.Secondly,we introduce the MBconv module as the decoder,enhancing the feature extraction capability.We then replace the SE attention module with the CA attention module to retain more spatial position information.Finally,we substitute the commonly used Batch Normalization(BN)layer with the Filter Response Normalization(FRN)layer to prevent model performance from being affected by batch size.Experimental results based on FeM and Cu datasets demonstrate that our proposed model achieves a PA of 96.58%and 95.39%,an MIoU of 92.8%and 90.43%,and an F1score of 95.15%and 93.47%.Compared to Unet,the Efficient_Unet model parameters are reduced by 60.4%,and the inference speed is improved by 19.23%,reaching 21.7frames per second.Our proposed model outperforms existing classical segmentation models in terms of accuracy and speed,exhibiting strong generalization performance.
作者
曾中华
曹东
Zeng Zhonghua;Cao Dong(School of Physics and Electronic Information,Henan Polytechnic University,Jiaozuo 454000,China;Wuxi Dongru Technology Co.,Ltd.,Institute of Industrial Intelligence,Wuxi 214000,China)
出处
《电子测量技术》
北大核心
2023年第21期176-182,共7页
Electronic Measurement Technology
