| With the vigorous development of multimedia information processing technology,the application of multimedia information processing has become more and more extensive in people’s lives.Digital image is an important form of multimedia information,and its related research and processing occupy an important position in many areas of life.Image classification belongs to the basic research content of digital image processing and has very important research value.Many related researches are carried out on the basis of image classification,such as target detection,image segmentation,and target recognition.Convolutional neural network model is the most studied image processing method in digital image processing in recent years.Compared with traditional classification methods,it has higher classification accuracy and better performance.There are also many shortcomings and difficulties that need to be overcome.This article has conducted in-depth research and improvement on the deficiencies of the widely used AlexNet network model and VGG-16 network model in image classification.Firstly,this paper introduced the basic structure of convolutional neural network in detail from the perspective of objective function optimization,conducts in-depth research and analysis on the basic structure of convolutional neural network,and a detailed derivation on the related calculation involved in convolutional neural network was gived from the perspective of objective function optimization.Secondly,for the problem of low classification and recognition accuracy of AlexNet convolutional neural network on multi-label images,the convolutional layer of the network was designed as a three-branch structure based on the AlexNet neural network.The ELU activation function was used which is more sufficient to extract features.The batch normalization operation was add to speed up the network learning rate.The improved network optimized on the Caltech-256 data set through the gradient descent algorithm.The result of experiments verified the effectiveness of the network improvement strategy.Finally,in view of the difficulty of training multi-label image classification by VGG-16 convolutional neural network and the relatively poor classification results,the VGG-16 convolutional neural network was improved by changing the convolutional layer connection mechanism and using the three-branch structure to change the network feature extraction method.The improved convolutional neural network was trained and tested on the UCM multi-label data set.Comparative experiments verified the effectiveness of the improved network.The improved network accuracy rate has increased by 15%,the recall rate has increased by 17%,and the F1 value has increased by 16%. |
