Research And Application Of Weight Initialization Method For Convolution Kernel In Convolution Neural Network

Convolution kernel is the main component of convolution neural network to extract image features.It detects the similarity between adjacent pixels by convolution calculation in the local area of the sample,so as to achieve the purpose of identifying image content.Therefore,the convolution core is an important factor affecting the recognition accuracy of convolution neural network,and it is also one of the reasons for the fast convergence of the network.How to initialize the convolution kernel has become a key problem to improve the performance of the convolution neural network model.For this reason,the following research is carried out in this paper:(1)The structure of convolution neural network,the process of image convolution and the process of feature extraction from convolution core are analyzed.Three commonly used initialization methods are studied:random initialization method,Xavier initialization method and Principle Component Analysis(PCA).Aiming at the problems of these three initialization methods,a method of initializing the convolution kernel weights by Kernel Principle Component Analysis(KPCA)is proposed.KPCA can effectively extract the non-linear features in the image,so that the weight of the convolution core can contain more non-linear features in the sample,so that the convolution kernel can be initialized more fully.(2)In order to verify the KPCA initialization method proposed in this paper,MN1ST,Fashion MNIST and Cifar10 are selected for training test.Firstly,the original data set is tested,and then the translation,rotation,elastic distortion and brightness change of the original data set are carried out The experimental results show that the proposed method can effectively initialize the convolution kernel weight of convolution neural network and improve the network recognition performance.Compared with PCA initialization,random number initialization and Xavire initialization,it has higher classification accuracy,reduces the loss value of the network,speeds up the convergence speed of the network,and solves the instability problem existing in traditional random methods.(3)KPCA initialization convolution kerne]method is applied to face recognition.OlivettiFaces is selected as the sample set for experiment.In order to verify the robustness of the algorithm,random horizontal flip,brightness change,image rotation,image movement and clipping operations are carried out on the sample set.Experiments on the expanded sample set show that the recognition accuracy of this method is better than the other three initialization methods,which verified that the method in this paper has certain applicability.