Compact Clarity: Development of Lightweight Networks for Low-Light Enhancement via Multi-Objective Knowledge Distillation

Low-light image enhancement remains a challenging task, often requiring large, computationally expensive models to achieve high fidelity. Existing state-of-the-art methods like MIRNet-v2 excel in structural restoration but can have high latency and significant parameters, while lighter models like Zero-DCE efficiently correct illumination but may lack detail preservation. To address this trade-off, we propose an efficient student network trained via progressive multi-teacher knowledge distillation. Our framework leverages the complementary strengths of MIRNet-v2 (structure, 26.6M parameters) and Zero-DCE (illumination) as teachers. Key innovations include: (1) Resolution-progressive training that adapts image size (64px→256px) and dynamically adjusts batch size for stable optimization. (2) A comprehensive hybrid loss function incorporating pixel, perceptual, frequency, structural (MS-SSIM), gradient, adversarial, and contrastive feature distillation terms. (3) A lightweight student architecture featuring Recursive Residual Groups (RRGs), Multi-scale Residual Blocks (MRBs), and optional Adaptive Feature Stretch (AFS) and Gradient-Guided Convolution (GGC) blocks for enhanced feature representation. Extensive experiments on the LOL dataset demonstrate that our model's generator, utilizing only 9.99M parameters (12.76M total including discriminator), achieves a PSNR of 21.89 dB and SSIM of 0.86. This represents a 2.66x parameter reduction for the generator compared to the 26.6M MIRNet-v2 teacher, while retaining 92.3% of its PSNR performance. Our approach enables efficient, high-quality low-light enhancement suitable for resource-constrained environments.