Research On Clothing Attribute Recognition And Landmark Detection Algorithm Based On Deep Learning

With the continuous improvement of computer performance and the rapid development of the computer visual,visual fashion analysis has drown lots of attentions.Recently,extensive research efforts have been devoted to clothes recognition,clothing item retrieval,fashion trend prediction and fashion recommendation.However,the study still confronted with those challenges as follow:(1)Basic problem in the field of visual fashion,the acquisition of clothing attributes mainly depends on manual annotation and maintenance.(2)It is difficult to eliminate the problem that the deformation of the scales presented in the clothing items affect the recognition performance,such as the distance and angle of camera shooting,and even how the apparel is displayed or the model is posing.We found that there is currently a lack of specialized and practical dataset in the field of visual fashion.This paper introduces a dataset ed by professionals and accord with the requirements of machine learning,and this study focuses on the clothing attributes recognition and the clothing landmark detection based on the dataset.Based on the existing work,this paper has carried out the following research:(1)To solve the problem that the attributes of clothing are acquired by manual annotation and maintenance,we proposed a convolutional neural network based on deep transfer learning and feature enhanced.In the feature enhancement branch,we innovatively propose an attention proposal network based on weak supervised learning to extract candidate regions.Firstly,the original and enhance branches are trained on the original images and the attention regions respectively.And then the multi-scale training of the network is completed by combining the original and enhanced branches,which makes the network more robust.Moreover,in the network training optimization,an activation function based on weight attenuation and separate multi-task joint training are added to improve the accuracy of clothing attribute recognition.(2)In order to solve the problem of clothing deformation and complex pose affecting the accuracy of recognition,a clothing landmark detection model based on improved cascaded pyramid network is proposed.The cascaded pyramid network combines the high-resolution of the shallow feature and high-semantic information of the deep features,and fuses the features of different layers to achieve more accurate prediction.We introduce the dilated convolution instead of the down-sampling operation in the model to reduce the normalization error of the clothing landmark prediction and improve the resolution of the feature map and increase the receptive field of the model.Then,an effective post-processing operation is introduced to further improve the performance of the model.Finally,the experimental results show that our model is effective,the relevant applications show the importance of landmark detection and explore the possibility of our research in the field of visual fashion.This paper hope to promote the development of related research fields by the research of clothing attribute recognition and clothing landmark detection,such as the alignment of fashion,the identification of local attributes and recommendation of clothing,the auto-editing and retrieval of clothing images,the navigation of tags and other application scenarios.