Research On Image Classification Algorithm Based On Convolutional Neural Network

In recent years,as an emerging machine learning technology,deep learning has attracted people’s attention for its excellent performance in image feature learning,among which deep learning models represented by convolutional neural networks have commendable results in image recognition and classification.Although the convolutional neural network can extract the deeper semantic information of the image well,it cannot learn independently,that is,it cannot pay attention to key information and ignore useless information.In addition,in practical applications,it is impossible to obtain a large number of sample data,which is the small sample problem,but the deep learning model represented by convolutional neural networks needs a large number of samples to train to obtain good performance,and cannot effectively solve the small sample problem.Based on this,this paper uses the convolutional neural network model as the theoretical basis to carry out the following research work:(1)The attention mechanism is introduced.In order to let the model learn to focus on key information and ignore useless information,so that image classification is more accurate and efficient,this paper introduces the attention mechanism.The traditional approach is one-way attention,and in order to better use the attention mechanism,combined with the spatial and background information of the channel,the bidirectional attention mechanism module(BDA)is designed,the two-way attention mechanism module can not only maintain the key information of the channel,but also maintain the spatial background information,can more accurately determine the importance of each channel,let the neural network focus on a specific channel,increase the help to the current task,reduce the negative impact on the current task,This makes image classification more accurate and efficient.A large number of experiments are carried out by integrating BDA with the three most classical ResNet,Inception and VGG models,and the results show that the bidirectional attention mechanism module can effectively improve the performance of the network model.(2)Aiming at the problem of small-sample image classification,this paper proposes a strategy combining active data enhancement learning and parallel fully connected network classifier.Firstly,the ResNet,Inception,and VGG network models of BDA are transferred through transfer learning,and then the active data enhancement learning strategy can effectively alleviate the problem of insufficient sample size in the training stage,and finally the parallel fully connected network classifier improves the performance of the classifier in the classification process.These three links complement each other,so that deep neural networks can also effectively classify small sample images.Finally,the experimental results show that the proposed strategy is feasible and effective on small sample datasets,which provides a research idea for the problem of small sample classification.