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

Clothing image segmentation is widely used as a basic function in the fields of virtual fitting,clothing matching recommendation,and clothing retrieval.Currently,deep learning methods are used to segment clothing images with rich colors,shapes,and complex backgrounds with good results,but there are still problems such as mixed segmentation of various parts in clothing images and inaccurate segmentation of clothing boundaries.In this paper,the following work is carried out to improve the quality of segmentation of clothing images by proposing a clothing image segmentation network,FMBNet:(1)A feature alignment module is designed for the feature unalignment problem that occurs in the fusion of semantic feature maps at different scales.By initially fusing different scale feature maps,the position reference of the high-resolution feature map relative to the lowresolution feature map is obtained,and the low-resolution feature map is upsampled according to the position reference and fused with the high-resolution feature map to obtain the featurealigned fused features.(2)A category attention mechanism is designed for the problem of segmentation errors in each part of the clothing.And the spatial,channel,and category attention are formed in parallel into a multi-directional attention mechanism module,which learns multi-directional semantic information from feature maps aligned and fused at multiple scales to obtain more detailed semantic information.(3)A boundary enhancement module is designed for the problem of inaccurate boundary segmentation of each part of the clothing.The boundary annotation map is generated by onehot encoding and distance calculation of the annotated image,and the boundary feature map is generated by the Laplace convolution operation on the feature map.And introduce the boundary supervision loss function to supervise the boundary feature map with the boundary annotation map.The proposed clothing image segmentation network FMBNet achieves higher quality segmentation of clothing images by improving the Mean Intersection Over Union on the Deepfashion2 dataset by more than 2%,5%,and 10% compared to OCRNet,Deeplabv3+,and PSPNet,respectively.In addition,to simplify the process of clothing image segmentation,a clothing image segmentation system is developed based on the segmentation algorithm and training weight file proposed in this paper,which can visualize and save the segmentation results,provide samples for future succession training.