Research On Landmark Detection And Classification Algorithm Of Clothing Image Based On Convolutional Neural Network

In recent years,online shopping has become an important channel for modern Chinese to purchase most of the goods,especially clothing products.Various e-commerce platforms are also committed to creating more convenient and comfortable online for users Shopping.People often search for clothing by entering category and attribute keywords on e-commerce platforms.However,there are numerous clothing on the Internet,and manual labeling of clothing categories and attributes is not only inefficient,but also subjective and error-prone.Therefore,how to use computers to efficiently and accurately label clothing with categories and attributes has become the main research direction of researchers in this field.In addition,the positioning of landmark of clothing is the basis of clothing category classification and attribute prediction,and it is also a significant research direction in the domain of clothing image.Today,deep learning algorithms are becoming more and more prevalent in image processing.Researchers also use convolutional neural networks to handle the tasks of clothing landmark detection and clothing classification.However,due to the easy deformation of the main clothing,the complex background and the occlusion of landmark of the clothing,the accuracy of the positioning and classification results still needs to be improved.Therefore,this paper uses deep learning algorithms to conduct in-depth research on clothing image landmark detection,clothing category classification and attribute prediction,the main work is as follows:(1)This paper designs an end-to-end deep convolutional neural network based on VGG16 to solve the three problems of clothing image landmark detection,clothing category classification and attribute prediction.(2)Aiming at the task of clothing landmark positioning,this paper designs a landmark detection method based on the predicted Gaussian heat map.First use the first four convolutional layers of VGG16 to extract features,and then add a global information extraction module,a deconvolution module and a multi-scale fusion module to the network,and finally generate a landmark heat map with the same size as the original image,and then get the landmark position.In addition,this paper uses the heat map obtained by the landmark detection network to generate the landmark attention,which lays the foundation for the subsequent clothing classification and attribute prediction tasks.(3)Aiming at the task of clothing category classification and attribute prediction,this paper designs a feature extraction method based on landmark attention and channel attention.First,use the landmark heat map to generate the landmark attention,and merge it into the feature map extracted by the first four layers of the VGG16 network,which improves the feature expression of the clothing landmark region in the feature map.In addition,considering the different effects of different channel features of the feature map on classification,a channel attention module is appended to the network to increase the weight value of important channels and reduce the weight value of noise channels,thereby suppressing the impact of noise on the classification process.(4)For the above three tasks,this subject is verified on a large-scale Deep Fashion data set.The experimental results prove that the clothing image landmark detection and classification algorithm designed in this paper has superior performance.Compared with previous studies,the accuracy of landmark detection has been significantly improved.In category classification,Top3 accuracy rate reached 91.24%,and Top5 accuracy rate reached95.95%.