Research On Deep Learning Algorithm Of Image Recognition And Object Detection

In recent years,deep learning technology has made a breakthrough in the field of computer vision,including two important basic issues: image recognition and object detection.Among them,image recognition is mainly applied in face recognition,finance and medical treatment in security field.Image detection is widely used in intelligent driving,video surveillance,industrial detection and many other fields.Image recognition and detection technology can reduce labor cost in many fields and has important practical significance.The image feature information will be lost in the convolutional layer or the full connection layer of the neural network,which can be effectively solved by the residual network.However,when the network is deep only a few units can effectively and fully extract features.At the same time,when depthwise separable convolution is used to feature extraction,the information between channels cannot be fused effectively,so that the feature map contains less semantic information.In view of the above problems,this paper proposes the method of channel cross fusion and constructs different model structures respectively for offline handwritten Chinese character recognition,face recognition and object detection.The specific work is as follows:1.A deep residual network model with multi-channel cross fusion is designed,which can make better use of the characteristics of residual network to extract features and improve the recognition accuracy.The improved center loss function is combined with the traditional Softmax loss function to improve the classification performance of the model.Experiments on CASIAHWDB-V1.1 data set show that the recognition model and algorithm designed in this paper can effectively improve the recognition rate of handwritten Chinese characters.2.To solve the problem of calculation of model of mass and takes up lots of memory resources,unable to meet the requirements of real-time and resource constraints,design two kinds of lightweight recursive residuals neural network and application on MobilenetV2 backbone network respectively,at the same time a gradient weighted global average pooling method is designed.The test results on the LFW,AgeDB-30 and CF data sets show that the network model designed in this paper achieves high face recognition accuracy with a large number of parameters reduced.3.Aiming at the problem that SSD target detection algorithm has poor detection effect on small targets,predecessors made many improvements,such as FPN,FSSD and other networks,but these networks did not consider the communication of information between feature map channels.Therefore,this paper designs a feature cross fusion model and a learning rate of periodic oscillation attenuation,which can effectively promote the information fusion between different channels.This learning rate can solve the problem of easily falling into the local optimal solution when training to a certain extent,and the experiment results are better.