Phase Contour Enhancement Network for Clothing Parsing

Clothing parsing is a challenging task. In this task, there are two main inherent challenges, which are the fine-grained parsing of small clothing components and the missegmentation of similar clothing categories. To address these challenges, this paper proposes a phase contour enhancement network for clothing parsing. It achieves improved accuracy while maintaining a relatively small network complexity. Leveraging the distinctive features of phase maps, we introduce the phase contour enhancement attention (PCEA) module to augment the encoder's object edge information. To further enrich the model's feature extraction capacity, we present the dilated convolution pyramid (DCP) module, combining it with a lightweight decoder to achieve improved global context modeling. The synergistic integration of the PCEA and DCP modules enhances the network with remarkable capabilities in capturing intricate contours, effectively surmounting the challenges of fine-grained parsing for small clothing components and mitigating the missegmentation of similar clothing categories. Through extensive experiments on the CFP, Modanet, and DeepFashion2 datasets and further expansion experiments on the LIP dataset, our method demonstrates outstanding performance across multiple evaluation metrics, surpassing state-of-the-art methods for clothing parsing. In conclusion, our phase contour enhancement network exhibits remarkable performance in both clothing parsing tasks and extended human parsing tasks.