Research And Application Implementation Of Clothing Segmentation Algorithm Based On Deep Learning

In recent years,the clothing fashion industry has developed rapidly,and people mainly obtain clothing information through images.Therefore,using the technology in the field of artificial intelligence to analyze and process clothing images has become one of the research focuses in the field of clothing.Among them,the image segmentation technology can divide the clothing image into different clothing areas,so as to facilitate subsequent analysis and processing.Currently,most of the methods belong to traditional segmentation methods or methods based on deep convolutional neural network(DCNN).However,these methods still have the problems of low fit of clothing region segmentation and poor clothing edge segmentation.Aiming at these problems,this paper proposes an improved method based on Deeplab v3+network,which adds object context information extraction module and edge optimization module.The object context information extraction module focuses on the context information aggregation method in semantic segmentation,and enhances the pixel feature representation with the help of the feature information of the corresponding clothing area,thereby improving the label classification effect of the pixel points in the clothing image and improving the segmentation fit of the clothing area.The object context information module mentioned in this paper is mainly realized through three parts of work:(1)Extracting the clothing image context is divided into a set of object regions,each region is a class,that is,using a deep network(for example,Res Net or HRNet)Segmentation to obtain rough segmentation results;(2)Calculate the similarity between the pixel point and the clothing area to obtain the clothing context information;(3)Use the clothing context information to enhance the pixel feature representation.The edge optimization module uses the classification labels of the internal pixels of the clothing image segmentation results to replace the classification labels of the original unreliable boundary pixels,so as to improve the edge quality of the segmentation results generated by any segmentation model,and greatly improve the edge segmentation effect.The edge optimization module consists of four parts:(1)Perform edge detection through supervised learning of the true edge map to obtain a good edge prediction map;(2)Find out the direction of the internal pixel associated with each edge pixel and generate the direction Prediction map;(3)Use a vector to represent the corresponding relationship between each edge pixel and its corresponding internal pixel.This vector starts from the edge pixel and associates an internal pixel to generate a coordinate offset map;(4)According to the offset map,the Garment segmentation results are subjected to edge refinement.Experiments have proved that this method has excellent performance on the Deep Fashion-Multi Modal dataset.Its m IOU index reached 67.29%,and its m Acc index reached 98.76%.The final results prove that our method is better than the original Deeplabv3+and its related methods.Optimize the network.The importance of this improved method lies in the fact that clothing image segmentation is the basis for obtaining clothing information,and accurate segmentation results can improve the subsequent analysis and processing of clothing images.The method in this paper uses the object context information extraction module and the edge optimization module to effectively solve the problems existing in the traditional method,improve the fit degree of clothing region segmentation and the effect of edge segmentation,and has better practicability and application prospects.In the future,this improved method can be applied to a wider range of clothing image analysis and processing scenarios to help people obtain and use clothing information more accurately.In addition,this improved method has certain universality and can be applied to image segmentation processing in other fields.By extracting and utilizing the object context information in the image,the accuracy and stability of image segmentation can be effectively improved,and it has great research and application potential.