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

With the rapid development of Internet technology,image information occupies a significant proportion in the explosive growth of various types of data information resources.It is important to know how to organize,analyze,and get the information what we need when we face with massive amounts of data.In real life,image information mostly contains rich semantic information,such as multiple targets,scenes,and behaviors.Therefore,the research on multi-label image classification method has more practical significance.The development of deep learning has brought people's research in the field of computer vision into a new era.Convolutional neural networks as a typical model in deep learning which have been successfully used in various fields of computer vision.In this dissertation,we analyze and deeply study various convolutional neural network models based on deep learning theory,and then propose our multi-label image classification methods based on these models.The final effectiveness is verified by experiments.The main research contents of the full text are as follows:1.This dissertation first expounds the research background,significance and research status of multi-label image classification methods and deep learning.Secondly,we have carried out in-depth research and analysis on multi-label image classification research tasks,including the traditional multi-label image classification methods and the basic theory of multi-label image classification method based on deep learning.At the same time,this thesis introduces the basic structure and ideas of the multi-label image classification method in detail,and analyzes the typical classification model.2.Aiming at the difficulty of capturing semantic association information in multi-label image classification method,we propose a multi-label image classification method based on convolutional neural network and fusion of attention mechanism and semantic relevance.Firstly,we use convolutional neural network to extract features.Secondly,the attention mechanism is used to correspond each label category in the data set and each channel in the output feature graph.Then,we use supervised learning to learn the relevance between channels.Finally,the experiments on MIRFlickr25 K and PASCAL VOC2007 datasets show the effectiveness compared with other methods.3.In this thesis,the problem of how to better model the relation of the network among labels is discussed.Inspired by the feature association module proposed in the field of target detection,the relevance of image features can be well learned.By designing a special network structure,the relevance among labels can be directly captured.Firstly,we integrate the dilated convolution into Resnet network,which enlarges the sensing field and improves the feature extraction information without using pooling operation to reduce the loss information.Then,we use the attention mechanism to classify the label features on the convolution feature channel dimension.Finally,we use the feature association module to model the correlation of the semantic features of the labels in the images to improve the multi-label image classification effect.