Multi-task Learning Based Skin Segmentation

Skin segmentation is a critical task in computer vision that has diverse applications in several fields such as biometrics, medical imaging, and video surveillance. Despite its importance, the acquisition of high-quality data remains a significant challenge in skin segmentation research. In this paper, we propose a novel skin segmentation algorithm for single-person images by utilizing a dual-task neural network built on the multi-task learning framework. Specifically, the algorithm employs an encoder-decoder architecture consisting of a shared backbone, two dynamic encoders, and a decoder. The dynamic encoders use dynamic convolution to extract more spatial location information, while the decoder utilizes a query-based dual-task approach that allows each task to utilize the information generated by the other one efficiently. The experimental results indicate that the proposed skin segmentation algorithm outperforms or matches the current state-of-the-art techniques on the benchmark test set.