Enhanced Hybrid Skin Detection Method Using Multiple 3D Volumes and Threshold Optimization

Accurate skin detection is vital in various applications of computer vision, including medical imaging, human–computer interaction, and facial recognition. Yet, delineating non-skin pixels from skin ones remains challenging due to diverse skin tones and lighting conditions. To this purpose, we introduce an innovative hybrid skin detection approach in HSV and YCbCr color spaces. The proposed method utilizes multiple 3D volumes to represent skin colors’ distribution. It employs Density-Based Spatial Clustering of Applications with Noise (DBSCAN) [1] to cluster skin colors’ distribution. For each cluster, we create statistical chrominance models at different luminance levels, as color distribution is not luminance-invariant. B-spline curve fitting generates 3D volumes by interpolating these models. Gradient descent optimizes the B-spline parameters for improved skin color distribution capture. Our method was evaluated using the SFA dataset and demonstrated highly satisfying qualitative and quantitative results against other hybrid and rule-based methods.