Research On Image Classification Algorithm Based On Attention Mechanism

In the context of the era of big data,massive amounts of image data are emerging in major search engines,and people can only use specific image classification algorithms to improve the search efficiency of image data.Image classification is one of the key research directions in computer vision tasks,aiming to correspond a given image to its category.At present,although convolutional neural networks have produced many representative network models,the performance of existing network models still cannot meet people's needs for classification accuracy,so the study of image classification tasks is still of great significance.This paper mainly studies the image classification algorithm based on the attention mechanism.The specific research contents are as follows:(1)For the study of image classification algorithms,the principle and advantages of convolutional neural networks are first introduced,and several typical network models in recent years are described.Then it introduces the attention mechanism in detail,and introduces a new attention visualization method to facilitate research and analysis.(2)Aiming at the problem that the residual network has insufficient learning of multi-scale image features and is difficult to be used in complex image classification tasks,an image classification algorithm based on dual attention module is proposed.The algorithm uses residual network as the basic network model.Improvements are made by fusing Pyramid Feature Attention(PFA)module and High-level Feature Enhancement(HFE)module to obtain a network model with better performance.The Pyramid Feature Attention module implements high-level features to guide the selection of low-level features,while the High-level Feature Enhancement module is a feature enhancement module that combines channel attention and spatial attention.The experimental results obtained on some ImageNet2012 datasets and Place2 datasets show that the improved network model has higher classification accuracy.(3)Aiming at the problem that the existing neural network model is difficult to learn the discriminative features of the target object in the image,a general lightweight group-wise attention module(LGAM)is proposed for image classification tasks.The module is decoupled along the channel and space of the input feature map.First,in order to ensure the lightweight of the module,the input feature map is grouped using the idea of grouping convolution,and the channel attention weight is generated on each group.At the same time,a ladder-shaped structure is adopted to alleviate the problem of information flow between groups;then,the results generated by each group are concatenated into a new feature map and global spatial attention weights are generated;finally,the fusion of the reconstructed feature map and the input feature map is enhanced Feature map.The experimental results obtained on the Cifar datasets and some ImageNet2012 datasets show that the combination of LGAM and existing neural network models can improve the accuracy of image classification.