Research On Fine-grained Image Classification Algorithm Based On Attention Guidance

Image classification is the most popular application direction of artificial intelligence in modern industry.Compared with general image classification,fine-grained image classification has the characteristics of large differences within classes and small differences between classes.Compared with the manpower cost of manual labeling with strong supervision,weakly supervised fine-grained image classification requiring only image labeling has become the mainstream research direction in the academic world.And the main research direction can be divided into CNN-based component localization and feature fusion method.Component localization focuses on mining the image region with the most discriminative force,and then making full use of the features in the region for fine-grained image classification.Feature fusion method focuses on end-to-end feature extraction of images.By comparing and analyzing the classical network model,this paper designs and implements a more excellent fine-grained image classification algorithm model.The main work of this paper can be summarized as the following three aspects:(1)Research on classical algorithms for fine-grained image classification.This chapter introduces NTS-Net algorithm based on component location,HBP algorithm based on feature fusion and MA-CNN algorithm based on visual attention respectively.This paper introduces the basic framework principle and model implementation details of each representative algorithm in detail,analyzes the advantages and disadvantages of the representative algorithm,compares and analyzes the accuracy of different algorithms combined with the commonly used fine-grained image classification data set,and finally analyzes and summarizes how to extract discriminant features.(2)Fine-grained image classification algorithm based on attention mechanism and graph convolutional neural network.Based on the importance of features,CBAM module is embedded into ResNet-50 backbone network to extract the most important feature regions of input images.In order to consider both global image-level background and local spatial background in the learning process of region recognition ability,local spatial context information is encoded by iterative cross-aggregation operation.Finally,graph convolutional neural network is added to explore the internal semantic association between feature vectors in order to obtain better recognition ability.Finally,the rationality and effectiveness of the proposed algorithm are verified by ablation experiment and comparison experiment.(3)Fine-grained image classification algorithm based on Transformer.In this chapter,the advantages and disadvantages of the algorithms in Chapter 3 are analyzed.Considering that visual Transformer has a self-attention mechanism that can extract global features of images better than ResNet-50,this chapter proposes a fine-grained image classification algorithm based on feature fusion Transformer,which can be roughly divided into: Global Transformer module,local Transformer module,multi-level feature fusion and result output.The attentional interaction module proposed in this chapter selects hidden features from each layer,replacing the input from the last Transformer layer with the selected tag.In this way,the class tokens in the Transformer layer at the next level fully interact with the low-level,intermediate,and advanced features at the next level,enriching local information and feature presentation capabilities.Finally,the rationality and effectiveness of the proposed algorithm are verified by ablation experiment and comparison experiment.