GAD-VAE: generative adversarial disentanglement with variational autoencoders for hair removal in dermoscopy images

This paper leverages the effectiveness of variational autoencoders in removing hair in dermoscopy images and presents an enhanced approach that combines variational autoencoders with generative adversarial learning and disentangled representation referred to as GAD-VAE to produce high-quality dermoscopy images that are free from hair. The proposed GAD-VAE employs disentangled representation to design a two-branch network for the generator. This network learns latent representations that distinguish hair-related features from non-hair-related features. Each branch is dedicated to a specific task. The first branch concentrates on generating hair-free images using variational autoencoders, while the second branch models hair characteristics. Notably, this model exhibits contextual awareness, explicitly capturing hair properties, thereby benefiting the second branch. The model also incorporates two discriminators that assess the realism of the generated hair-free images and hair images. Experimental results demonstrate the potential of the proposed model in effectively removing some types of hair compared to variational autoencoders and some of the state-of-the-art methods. Additionally, it generates hair that can be utilized to develop hair segmentation methods.