DASP: Dual-autoencoder Architecture for Skin Prediction

This paper introduces a novel human skin detection approach based on the application of a dual autoencoder architecture, composed of models to detect background and skin zones concomitantly. This method, named Dual-Autoencoder Skin Predictor (DASP), associates the outputs of two autoencoders through a composite loss that minimizes the error between predicted skin/background areas and the groundtruth. More importantly, the composite loss penalizes overlapping zones between autoencoders predictions, leading our approach to better capture fine-grained and complementary information between skin and background. To combine semantic information with the skin color distribution, heavily tackled by handcrafted skin detection methods, our architecture relies on a main input that considers multiple colorspaces. Besides, a secondary input provides a standard skin/background patch vector to the model, granting information regarding their color distribution. Our experiments support the accurate performance of the proposed architecture and highlights the contributions of the composite loss and multiple inputs. For instance, DASP achieves the best and the second best results on Pratheepan and Mutual Guidance datasets, respectively.