Research On Image Recognition Based On Optimized Convolutional Neural Network

In recent years,the study of deep learning technology has aroused widespread concern in academia and industry,and has promoted the rapid development of artificial intelligence.As an important branch of deep learning research,convolutional neural network has made breakthroughs in the field of pattern recognition and image processing because of its three characteristics of local receptive field,weight sharing and downsampling.However,there is still a lot of research space in terms of optimizing the network structure or processing massive data sets.The main research contents and innovative work of the paper are as follows:(1)In order to improve the accuracy of feature extraction by convolutional neural networks and improve the robustness of the model to three transformations of rotation,translation and scaling,an adaptive weighted pooling algorithm is proposed in this paper.The algorithm combines the advantages of maximum pooling and averaging pooling while incorporating adaptive weighting masks into the aggregation operations,which can indicate that the convolutional neural network is used to identify important features of the image.Experiments show that the adaptive weighted pooling algorithm is applicable to various image databases and network structures,which can significantly improve the classification effect of the network.(2)Aiming at the gradient disappearance problem and the negative interval neuron necrosis in the existing activation function,this paper analyzes the causes of these phenomena and proposes a parameter-adjustable activation function Fexp.This activation function can freely scale the gradient by adjusting the parameters appropriately,thus alleviating the gradient disappearance problem.Experiments on the MNIST handwritten database and disk dataset prove that the Fexp function isfeasible,and the best performance such as high accuracy and fast convergence is obtained.(3)In order to solve the problem of high dependence on a large number of labeled samples and high cost of mass data in the model training process,this paper proposes a sample optimization method based on active learning.This method uses information entropy and diversity as rules to effectively utilize the information contained in the image.In each iteration of the network training process,it is preferred to train a sample with a large amount of information and rich diversity,thus a higher classification accuracy can be achieved by using fewer training samples.This sample selection method greatly reduces the cost of manual labeling while taking into account the accuracy of the classifier.