Clothing Parsing Based On Attention Model

Clothing not only meets the needs of people's daily life to keep warm and get sheltered,but also reflects people's fashion sense,individual identity and status.Therefore,more and more works try to analyze clothing data,such as clothing retrieval,clothing recommendation,virtual try-on,etc.Clothing parsing is also brought into being.With wide application of deep learning technology in the field of computer vision,the researches in clothing parsing also have some achievements.However,in order to improve the analysis performance of the model,the current clothing parsing work often designs complex network structure or references additional auxiliary information,which leads to large parameters and complex calculation,so it does not fully meet the needs of practical application.Therefore,this paper deploys two relatively lightweight attention modules on the backbone network with a small number of parameters to solve the problem of poor discriminative representation due to inefficient feature fusion or lack of feature modulation and semantics dilution in deep decoder blocks when propagating information,so as to improve the parsing performance of the model.The specific works of this paper are as follows:Firstly,we propose a clothing parsing method based on gated mechanism attention model.At present,the most advanced clothing analysis methods often use encoder-decoder structure.However,in the fusion process,the feature information of shallow encoder blocks and deep decoder blocks are spliced without screening.To solve this problem,an attention model based on gated mechanism is designed.The attention model adaptively selects the missing context information that the deep decoder block feature information from the shallow encoder block,and uses this information to supplement and fuse the deep decoder block feature.Experiments show that this method can improve the parsing performance of the model,and does not significantly increase the size of the model.Secondly,we further propose a clothing parsing method based on unabridged attention and adjacent modulation.The proposed novel unabridged channel attention module can recalibrate the features in encoder blocks and decoder blocks,so that the refined features are more suitable for clothing parsing task.In order to solve the problem of semantic dilution between deep decoder blocks,the proposed top-down adjacent modulation module take semantics-consistency constraints on the feature information of adjacent decoder blocks(including the last encoder block and the first decoder block),which can gradually supplement high-level semantic information to each decoder block.Experimental results indicate that compared to the clothing parsing method based on gated attention model and other state-of-the-art clothing parsing methods,this method can achieves more competitive performance but with fewer parameters.Finally,we design and implement an image annotation system for clothing parsing task so as to solve the lack of high-precision clothing datasets.In this system,the image to be labeled is input into the trained attention-based clothing parsing model proposed in this paper to get preprocessing,then the user can further optimize the preprocessing results by the provided manual fine-tuning function.