Research On Image Glassification Algorithm Based On Convolutional Neural Network

In recent years,with the iteration update of hardware resources,deep learning algorithm has been used widely.Deep convolutional neural network is a very important network structure in deep learning.It simplifies the network model and improves the training efficiency through weight sharing,local connection and downsampling.Image classification is an important research direction in the field of computer vision.It refers to image processing technology that realizes different classifications by acquiring the salient features of images.Convolutional neural network realizes the convolution of image pixels and extracts image features directly from image pixels.This method is closer to the processing method of human brain visual system than the general feature extraction method,and the recognition effect is better.Not only limited to the image classification,convolutional neural networks also has made good progress in face recognition,target detection,speech recognition,intelligent monitoring and other directions,which has important research significance and value.This paper focuses on the influence of convolutional neural network structure on image classification results.A brief introduction to the deep learning classic network AlexNet,and based on it to design the network from different angles.From the convolutional layer extended to the parallel convolutional neural network,to the parallel scale cropping convolutional neural network,experiments were carried out on the database Caltech-101 and Caltech-256.By comparing and analyzing the changes of classification accuracy of the improved network on the two databases step by step,the better parallel network was determined.In addition,the batch normalization and feature fusion algorithms are introduced into the extended network,and the experimental accuracy is significantly improved.The main research work and innovations of this paper include:(1)Aiming at the problem of excessive parameters of parallel convolutional neural network(PCNN)and high training time cost,a parallel scale cropping convolution neural network(PSC-CNN)is designed by introducing the random crop layer of full convolution neural network into PCNN.Compared with the original parallel network,it has more random feature inputs and significantly reduces the computational burden of the network.The experimental results show that the improved network improves the classification accuracy and shortens the training time of the model.(2)The two-channel feature fusion convolutional neural network combined with batch normalization was designed by introducing batch normalization and feature fusion method.First,we design a DCNN,then introduce the Batch Normalization(BN)layer to construct convolutional unit block(Conv-Unit),and combine the idea of multi-level feature fusion to construct a multi-layer feature fusion convolutional neural network MFF-DCNN.On this basis,the idea of dual-channel multi-scale feature fusion(DCFF)is proposed.Analyzed the influence of different Batch size values on the recognition accuracy of network model.