Brightness-Aware Generative Adversarial Network for Low-Light Image Enhancement

Images from low-light frequently exhibit poor visibility, excessive noise, and color distortion, which substantially impair both computer vision systems and human visual perception. Although numerous enhancement techniques have been developed, producing visually appealing results with well-maintained structural details and natural color reproduction continues to pose significant challenges. To address these limitations, this paper present an Brightness-Aware Generative Adversarial Network (BA-GAN) for robust low-light image enhancement (LLIE). Our framework employs a U-Net-based generator that effectively captures multi-scale contextual features while preserving fine image details through skip connections. The key innovation lies in our novel Brightness Attention Mechanism Module, integrated within the decoder, which dynamically directs the network's focus to regions requiring substantial illumination correction. To ensure local photorealism, this paper adopt a PatchGAN discriminator architecture. The complete model is trained on the LOL dataset using a composite loss function combining: (1) adversarial loss for realistic image generation, (2) brightness attention loss for keeping the brightness accuracy, and (3) perceptual loss to maintain structural and semantic fidelity. Extensive experiments validate that our BA-GAN outperforms current state-of-the-art methods, achieving superior performance on both quantitative metrics (PSNR: 20.7127, SSIM: 0.7963, LPIPS: 0.2271) and qualitative visual assessments. The enhanced images demonstrate significantly improved visibility while effectively suppressing noise and preserving natural color characteristics.