Multi‐Level Adaptive Lighting Filter (MLALF): A Hybrid In‐Network Illumination Normalisation Module for Human Skin Detection

ABSTRACT

Illumination variation is a major source of degradation in human skin detection. Conventional preprocessing such as Contrast‐Limited Adaptive Histogram Equalisation (CLAHE) is optimised for perceptual quality and cannot be tuned to the segmentation objective, while feature‐space attention modules cannot correct extreme exposure at the input level. We introduce the Multi‐Level Adaptive Lighting Filter (MLALF), a lightweight, plug‐and‐play module placed before a U‐Net encoder that couples a deterministic, parameter‐free gain stage for global exposure correction with three trainable attention branches (global, local, and fine‐detail) optimised jointly with the segmentation loss. On four benchmark datasets, MLALF consistently improves segmentation over the U‐Net baseline and outperforms CLAHE, adaptive gamma correction, and the Squeeze‐and‐Excitation (SE), Efficient Channel Attention (ECA), and Convolutional Block Attention Module (CBAM) attention modules under an identical training protocol, with gains that are consistent across five random seeds and robust to small changes in the gain thresholds. Adding only about 0.21 M parameters (+0.73% computation), MLALF offers an effective illumination‐normalisation block for encoder‐decoder skin detection.