Aiming at the problems of the current real-time semantic segmentation algorithms based on clothing segmentation (such as Deeplab series, ERFNet series, BiseNet, etc. ), such as insufficient detail features, poor segmentation accuracy, slow reasoning speed, etc., a deep dual-resolution fusion multi-scale real-time semantic segmentation algorithm MDRNet was proposed. In order to improve the learning ability of clothing image features, a dual -resolution feature network was used to extract high-resolution and low-resolution features simultaneously, and a bilateral fusion module was used to integrate features of different resolutions for several times, thus improving the deep supervision output structure of the high-resolution network, and a cross-space multi-scale attention module was added to capture cross-space feature information. Finally, the context extraction module DAPPM was changed to the parallel module PAPPM, which not only improved the accuracy, but also effectively reduced the number of parameters and computational complexity, and effectively improved the FPS. The experimental results showed that MDRNet improved the average crossover ratio of DeeplabV3+and BiseNetv2 algorithms by 9.2% and 8.1% respectively on the DeepFashion-MultiModal dataset for clothing segmentation, providing a better technical solution for the application of semantic segmentation in clothing.