Research On Multi-label Image Classification Algorithm Based On Deep Learning

With the rapid development of computer and deep learning technology,image information occupies an increasing proportion of various data information resources.Faced with a large amount of data information,how people analyze,organize,and obtain the data information they need becomes particularly important.In real life,images usually contain rich semantic information,such as multiple objects,behaviors,scenes,etc.Therefore,it is of practical significance to study multi-label image classification methods.In this thesis,for the problem of multi-label image classification,based on the relevant theories and methods of deep learning,a variety of multi-label image classification algorithms are analyzed in detail.In particular,on the basis of the current popular multi-label image classification algorithm ML-GCN,two deep learning-based multi-label image classification methods are proposed.The main research content of the full text is as follows:(1)Aiming at the problem that the image features obtained by global maximum pooling in ML-GCN lack pertinence for specific categories on different image regions and lose image local feature information,a class-specific residual attention(CSRA)module is proposed.This module can effectively capture different spatial regions occupied by different categories of objects.Furthermore,by combining class-specific residual attention with graph convolutional neural networks,a multilabel image classification algorithm(ML-CSRA)based on multi-head class-specific residual attention and graph convolution is designed.Experimental results on MSCOCO 2014 and VOC-2007 datasets show that ML-CSRA outperforms existing algorithms in all evaluation metrics.(2)Aiming at the lack of flexibility of ML-GCN in modeling node correlation and ignoring the correlation of label node features,and the problem of weakening the similarity of original node features in the process of graph convolution feature aggregation,a method for multiple Adaptive Attention Graph for Labeled Image Classification is proposed.This graph introduces the correlation and similarity of tag features on the basis of tag co-occurrence probability to model the tag node relationship.In addition,a multi-label image classification algorithm(ML-AAGCN)based on Adaptive Attention Graph Convolutional Network is designed on the basis of Adaptive Attention Graph.Experiments were carried out on the MS-COCO 2014 and VOC-2007 data sets.ML-AAGCN achieved good experimental results.Compared with other algorithms,the strategy adopted by ML-AAGCN can improve the performance of multi-label image classification to a certain extent.