Automated method for hair removal in dermoscopic images using adaptive retinex

Résumé fourni par la source

The increasing occurrence of skin cancers necessitates accurate dermoscopic diagnosis. However, hair artifacts in dermoscopic images occlude critical lesion details, significantly lowering the accuracy of automated classification. This paper introduces a robust algorithm that employs the Adaptive Retinex (AR) theory to detect and remove both dark- and light-colored hair automatically. Our method first decomposes the image into the reflectance (ℛ) and illumination (ℒ) components (𝐼 = ℛ ⋅ ℒ). Hair segmentation primarily occurs on the reflectance map (ℛ), where a Difference of Medians (DoM) filter is applied to isolate hair shafts based on their directional morphology. An initial binary mask is generated using anisotropic filtering and double thresholding, then refined using luminance constraints from the illumination map (ℒ) for improved accuracy. Finally, exemplar-based inpainting synthesizes realistic skin textures to remove the detected hair smoothly. Tested on the HAM10000 dataset and compared it with six advanced techniques, quantitative metrics (PSNR, SSIM), and visual inspection demonstrate that our method more effectively eliminates hair while preserving the underlying texture. These superior results indicate that the proposed technique is highly suitable for subsequent lesion segmentation and feature analysis, thereby enhancing the diagnostic potential of automated dermoscopic systems.