Research On Object Detection Algorithm Based On Deep Learning

Object detection is a very important research direction in current computer vision tasks.That is to say,the position of the target is marked on the input picture,and the category of the object in the label box is determined at the same time.Object detection of the visual tasks has been rapidly developed in the current application of deep convolutional neural networks.Although new algorithms are emerging in the field,as well as more efficient feature extraction networks,there is still a big challenge in this field of detection,and this means that there is still much room for improvement in the visual task of target detection.This paper introduces the principle of target detection under deep learning,and analyzes some commonly used detection algorithms in detail.Finally,analyzing and improving the algorithm focused on the two-stage target detection,and improving and optimizing this algorithm to improve the detection rate.The main contents of this paper have the following three points.(1)Comparing and summarizing the related knowledge of target detection algorithms under deep learning.Introducing and comparing the internal modules(convolution layer,fully connected layer,pooling layer)and activation functions frequently used of deep convolution networks.At the same time,the classic network structure under deep learning is described and analyzed in detail.Finally,describing the two detection schemes of the object detection in detail: 1 a two-stage detection scheme based on regional proposal;2 a one-stage detection scheme based on regression.(2)A multi-layer feature fusion target detection algorithm based on RPN structure is proposed.The current objection detection algorithms usually only use the deep feature of the CNN feature extraction network to detect the object,and it can't make good use of shallow feature.In order to utilize the shallow feature to improve the richness of the finally extracted feature layer information,this paper presents a multi-feature fusion object detection algorithm based on RPN network structure,that gets different layers feature by deep convolutional neural network,and makes these shallow feature merge with these deep feature to improve the richness of the finally extracted feature layer information,to improve the performance of the detection model.(3)A human target detection algorithm based on improved VGG feature extraction network is proposed.The algorithm performs training detection on the human body in the actual scene,and improves the network framework of VGG.On the basis of the VGG extraction feature network,a 1*1 convolution and a maximum pooling layer are added to the first convolutional layer to change the feature dimension and the channel number,so that the feature dimensions of each subsequent layer are matched.Performing training on human target on the data set made by myself,the final detection effect is better than the unmodified two-stage VGG detection network.The final detection results show that the improved detection algorithm can better detect human targets.