Research Of Image Recognition Technology Based On Improved Convolutional Neural Network

Image recognition technology is one of the research hotspots in the field of pattern recognition.Its task is to use the prediction model learned from the training set to determine the category or attribute of a given image.In recent years,with the development of computer technology,the method of learning prediction models using Convolution Neural Network(CNN)technology has become the mainstream in the field of image recognition.Support Vector Machines(SVM)-driven CNN models use SVMs which based on large interval ideas as energy functions to guide the learning of CNN.Compared with traditional CNN models that use Softmax loss,the SVM-driven CNN have stronger generalization performance.However,However,the algorithm ignores the influence of the radius of the minimum enclosing ball(MEB)containing all samples in the feature space on the upper bound of the SVM generalization error,which greatly limits the further improvement of its generalization performance.In addition,In addition,the algorithm does not consider the distribution information of the overall sample,which making it difficult to guide CNN to extract image features with better quality.In view of the above two shortcomings,this article mainly did the following research:(1)During the training of the deep model,the feature space extracted by CNN is constantly changing.At this time,the upper bound of the generalization error of the SVM not only depends on the classification margin between different classes,but also the radius of the MEB in the feature space which is changing.In response to this fact,a strategy based on SVM generalized error bounds is used to guide the learning of CNN deep models and the construction of corresponding classifiers,then a CNN model driven by Radius Margin Bound(RMB)is proposed.Compared with the traditional CNN model,the proposed model can increase the classification margin between image features of different categories,and limit the increase of the MEB radius as much as possible,which ultimately drives CNN to extract higher quality image features.(2)SVM uses only a small number of support vector points in constructing the classification hyperplane,without considering the distribution information of the overall sample,and SVM-driven CNN will inevitably inherit this deficiency.Aiming at the insufficiency,a minimum class variance support vector machine(MCVSVM)combined with Fisher's linear discriminant theory is used to guide the learning of CNN deep models,and then proposed A CNN model driven by MCVSVM.The proposed model not only pays attention to the margin between categories during training,but also makes full use of the distribution of samples to fine-tune the classification hyperplane,thereby obtaining a classification hyperplane that is better than the SVM-driven CNN model to guide the learning of CNN.Finally,CNN can be driven to extract image features that contain more discriminative.In order to verify the effectiveness of the two proposed algorithms,three representative deep convolutional network structures include AlexNet,VGGNet,and ResNet were used.the feature visualization results and recognition accuracy of the proposed two algorithms are compared with Softmax-driven CNN,SVM-driven CNN and center-loss CNN in five large-scale data set FER2013,MNIST,SVHN,CIFAR-10 and CIFAR-100.moreover,the influence of parameters on model performance is also analyzed.Experimental results show that the two proposed algorithm has stronger feature expression ability and higher recognition accuracy than the competition algorithm.