An Adaptive Convolution Kernel-based Neural Network Algorithm

With the development of computer science and big data science,deep learning is widely used in many fields of artificial intelligence,such as image recognition,target de-tection,automatic driving etc.The convolutional neural network,that is the most popular deep learning model,can automatically extract the features of two-dimensional data in-stead of the complex artificial feature design in the traditional methods.It also can greatly reduce the model parameters compared with the traditional full connection method.This paper majorly studies an adaptive convolution kernel for convolutional neural networksFirst,a new adaptive convolution kernel is designed.The parameters of the tradi-tional convolution kernel are determined by the global feature information,so it is diffi-cult to extract the local feature information,which leads to the information fuzziness of the feature map.Therefore,we propose a new adaptive convolution kernel.Here,the kernel parameters depend on the local information of the feature map adaptively,which can improve the ability of the convolution kernel to extract the local details.Thus,it can effectively overcome the disadvantages of the traditional convolution kernelThen,the back propagation algorithm of network layer based on adaptive convolution kernel is derived.We consider the gray-scale image of 5 × 5 as the input,and take the adaptive convolution kernel with the size of 3 × 3 as an example,we give the gradient formula of the proposed adaptive convolution kernel parameters.Furthermore,the error back propagation formula of the adaptive convolution layer is derived in detailFinally,based on mnist handwritten data set,the validity of the adaptive convolu-tion neural network model is verified by comparing with the traditional convolution neural network model.Here,the influence of convolution kernel size,convolution step length and number of convolution kernels are studied respectively.And the results show that the present adaptive convolution network has a signifcant improvement compared with the traditional convolution neural network model in the model precision.Furthermore,the feature map of the convolution layer are studied.Comparing with the traditional convo-lution layer,we find that the feature map of adaptive convolution layer has a significantly better effect on the resolution of contour and edge.Finally,the adaptive convolution layer is successfully applied to the LeNet model.