| Summary: | With the development of online shopping platforms, consumers and designers need to choose from a large number of garments when shopping or designing. Quick identification of clothing products can effectively improve the efficiency of designers’ and consumers’ experience. Therefore, this paper used DeepLabV3+ combined with deep separable convolution to improve the network computation speed. To address the problem of low recognition rate of H-shaped silhouette in semantic segmentation, the fuzzy trouser silhouette samples are further analyzed. The trouser silhouette was redefined according to the characteristics of pants, and the dataset labels were optimized with a trouser silhouette classification method. It was found that the accuracy and efficiency of trouser silhouette recognition were significantly improved. The indicators of recall rate, IoU and PA of H silhouette is improved by 6%, 5%, and 1% respectively. After label optimization, the classification prediction accuracy of silhouette V is 100%, the recall of silhouette V is 97%, and the recall of silhouette O is 96%.
|
|---|