Research On Outfits Object Detection And Intelligent Composition Evaluation Based On Deep Learning

With the development of the Internet industry,e-commerce market ushered in unprecedented rapid development,which also led to the transformation of traditional industries represented by clothing from offline to online,online sales has become one of the important sales channels of the clothing industry Although there is no time and space constraint in shopping for clothing on the Internet,the inability to try on clothes can make consumers spend a lot of time and effort in selecting and composition,thus reducing the desire to consume.In order to improve the shopping efficiency of consumers and give outfit suggestions for them,this paper proposes a deep learning-based clothing detection network and a composition evaluation network,with the following main work:(1)An improved algorithm of YOLOv5 s incorporating attention mechanism is proposed to embed the attention mechanism before the spatial pyramid pooling of YOLOv5 s to reasonably allocate computational resources,reduce the interference of irrelevant information,and improve the efficiency of feature extraction.In this paper,we try to incorporate three attention mechanisms suitable for lightweight networks: SENet,CBAM and CA.m AP is improved by2.73% after embedding SENet,by 0.37% after embedding CBAM,and by 3.17% after embedding CA.(2)Based on YOLOv5s,we make lightweight improvements to the network and propose a lightweight network YOLOv5s-GC based on attention mechanism,which uses Ghost convolution to reduce the number of parameters of the network,uses Ghost Net V2 bottleneck to replace the original bottleneck residual block,and adds coordinate attention mechanism before the spatial pyramid pooling module.Experimental results on the Deep Fashion2 clothing dataset demonstrate that the YOLOv5s-GC network improves m AP by 2.5% with a 47.9% reduction in the amount of parameters computed by the network model(3)A multimodal outfit composition analysis network based on split-attention mechanism is proposed,which is based on a residual neural network stacked with split-attention mechanism,a multilayer comparison network and multimodal feature fusion to extract and model multimodal features from images and texts of clothing,to rate the overall composition and to give composition suggestions by comparing the composition degree of two pairs of clothing.(4)An intelligent detection and composition platform for clothing is developed based on Pyqt in Windows,which can detect and classify the clothing images in different scenes,and can also score the clothing composition and give composition suggestions.