Research On Clothing Detection Algorithm Based On Deep Learnin

In the current era of rapid development of e-commerce,clothing products not only account for a large proportion of people’s daily lives,but also are closely related to people’s lives.In order to improve users’ clothing shopping experience and comfort,various ecommerce platforms rely on computer vision related technology and the clothing field to continuously update and optimize the functions of e-commerce platforms.They have launched functions such as image search,clothing matching,and virtual fitting,The basic implementation behind these functions is clothing detection,and the effectiveness of clothing detection directly affects the presentation of results in upstream tasks of clothing.Based on the background of existing clothing detection models,this article applies deep learning models to clothing categories,taking into account the speed and accuracy of clothing networks.Based on this,further exploration of model lightweight is carried out,and ultimately,a lightweight and efficient clothing network model is constructed.The specific research work is as follows:(1)Explore and construct a clothing network architecture based on deep learning environment.In the case of insufficient Receptive field for large target detection,the accuracy of network detection is improved by deepening the depth of feature extraction network and adopting the method of four feature layer prediction;By comprehensively considering the network speed and accuracy,a structural reparameterization network is adopted as the feature extraction network.Different structures are used in the training and inference stages to balance the model speed and accuracy;Considering the importance of location information in feature maps,the spatial location information of feature maps is modeled,and a long-distance dependency based on location attention mechanism module is proposed to capture location information;In the process of feature pyramid fusion,set a set of adaptive feature fusion parameters for each scale feature map,enhance the richness of feature fusion,and introduce a loss function based on distance control,so as to avoid the problem of large computation in the derivation of inverse trigonometric function,thus accelerating the convergence speed of the model.(2)Lightweight design of clothing models.In view of the large number of parameters of the existing clothing detection model,which is difficult to deploy on mobile devices,the method of model pruning is used to prune the channels of the feature map: first,the clothing detection model is iteratively trained,then a scaling factor is introduced to each channel of the feature map,and an L1 regularization is applied to its factor for sparse training,and then different pruning rates are set to complete the channel clipping,And fine tune the pruned network.The method of channel pruning is used to compress the model.Compared to the original YOLOv5 s model,the FPS of the model has been improved,reducing the number of parameters by 64.7% and increasing the accuracy value by 2.16%.This combines the speed and accuracy of the model,achieving a lightweight design of the clothing detection model.