摘要
针对传统卷积神经网络层级较为浅,对物体识别精确度较低的原因,利用改进的深层卷积网络VGG16模型检测视频运动目标.首先,预处理过程中对数据集进行剪裁和旋转操作,补充数据集数量,以解决前期图像资源不足等问题;其次,在PASCAL VOC数据集上先预训练模型,接着加载自定义视频数据集对预训练模型进行第二次训练.实验结果表明,该网络模型能很好用于视频目标识别,提高了检测精确度,有效减少网络参数计算量,降低硬件内存资源消耗,具有较强的鲁棒性.
Because the object detection accuracy is relatively low for the traditional convolutional neural network,so the convolution network based on more deeper structure VGG16 is used to detect video moving object in this paper.First of all,during preprocessing,cropping and rotating operations of data sets is applied in order to solve the shortage of early image resources and other problems.Secondly,the pretraining model of VGG16 network is adopted by PASCAL VOC data set.Then the second training is based on the the custom video data set.The result shows that the network model can be effectivity applied to video object recognition.The model has the high detection accuracy,reduces the calculation of network parameter and the consumption of memory resources,and has strong robustness.
作者
杨洁
陈灵娜
林颖
陈宇韶
陈俊熹
YANG Jie 1,CHEN Lingna 1,LIN Ying 1,CHEN Yushao 1,CHEN Junxi 2＊(1.School of Computer,University of South China,Hengyang,Hunan 421001,China;2.Affiliated Nanhua Hospital,University of South China,Hengyang,Hunan 421002,Chin)
出处
《南华大学学报（自然科学版）》
2018年第4期61-68,共8页
Journal of University of South China：Science and Technology
基金
国家自然科学青年基金研究项目(61504055)
湖南省自然科学青年基金项目(2015JJ3110
2015JJ3105)
湖南省教育厅项目(15C1184)
湖南省研究生科研创新项目(CX2016B445)
关键词
卷积神经网络
SGD梯度下降
视频目标检测
模型训练
convolution neural network
SGD gradient descent
video target detection
model training
