Attention-Driven Multi-Scale GAN with Complementary Learning Sub-Network for Low-Light Image Enhancement

Low-Light image enhancement (LLIE) is a vital aspect of downstream vision applications like surveillance, photography, and autonomous systems, where inferior lighting conditions result in degraded image quality and impaired performance of such vision tasks. This research presents AM-GAN, an Attention-Driven Multi-Scale GAN, which integrates a Complementary Learning Sub-Network (CLSN) to address the challenges demonstrated by the non-uniform lighting. The CLSN generates the inverse grey maps as the attention maps, thereby guiding the enhancement process by emphasizing the under-illuminated areas while preserving well-illuminated regions. The model employs a U-Net-Based generator with multi-scale feature extraction from both the spatial and frequency domains, thereby attending to texture preservation and natural color consistency. A dual Markov discriminator evaluates both local (patch-based) and global consistency, ensuring the visual coherency of the generated results. The model is trained and evaluated in quantitative and qualitative aspects on multiple benchmark datasets, including LSRW, LOLv1, LOLv2-real, and LOLv2-synthetic. Results demonstrate the potential of the proposed method towards state-of-the-art methods with averages of 18.64 dB of PSNR and 0.468 of SSIM values while maintaining perceptual quality. Ablation studies demonstrate the significance of CLSN and the frequency feature integration. The proposed model is produced as a solution for enhancing low-light images, with potential application across various domain-specific visual data in challenging lighting conditions.