Feature Enhanced Deep Networks For Object Detection Algorithm

Computer vision refers to the realization of human visual functions through computers,so as to realize the perception,recognition and understanding of three-dimensional scenes in the objective world.In the early stage of the development of computer vision,the research direction of target detection has stimulated scholars' strong research interest.With the development of computer technology,the common target detection is to locate and recognize the target according to the characteristic information of different objects.Nowadays,face detection,behavior detection and other specific target detections have more mature technologies.In recent years,with the continuous development and maturity of deep learning technology,target detection technology based on deep learning has gradually become a research hotspot.At present,the target detection algorithm based on Convolutional Neural Network(CNN)has become the extensive used technical method in computer vision.In view of the excellent performance of deep learning theory and methods in target detection,this thesis proposes a two-stage target detection algorithm based on convolutional neural network,that is,deep target detection algorithm with enhanced features,aiming at the problem that the key content of the image cannot be effectively obtained due to pooling and other operations in the convolutional neural network.Specifically,the Faster R-CNN network cannot obtain the key perception domain of the target due to the use of shallow network structure(VGG16),and thus it is unable to extract feature maps with richer feature information.Therefore,given the issue of Faster R-CNN algorithm,this paper builds a deeper feature extraction network to improve the feature acquisition ability and integrates the hollow convolution operation into the deep feature extraction module based on its network structure.In this thesis,a large number of experiments were carried out on the standard database PASCAL VOC2007,and the influence of various parameters on the experimental results was thoroughly analyzed.The experimental results show that the network structure constructed in this thesis has good target object detection capabilities.The network can also effectively complete the classification task and obtain high recognition accuracy.Furthermore,the average accuracy of the methodproposed in this paper on target detection tasks is about 5% higher than Faster R-CNN.The remarkable performance indicates that this method has a broader application prospect.