CONTRAST-DRIVEN NETWORK FOR TWO-STAGE LOW-LIGHT IMAGE ENHANCEMENT

Abstract

Low-light image enhancement aims to convert poorly illuminated photos into visually pleasing, well-lit versions while preserving natural color and structural details. Although many classical and modern deep-learning methods successfully increase brightness, common issues persist: color casts, over/under enhancement, loss of texture, and dataset-specific behavior that limits generalization. This work proposes a self-contained contrast-driven encoder–decoder framework that removes dependency on external color-specific datasets by integrating two novel training principles: color constancy loss (a data-agnostic color balance prior) and perceptual loss (to preserve high level structure). The framework presented is designed for achieving consistent improvement in different lighting conditions without using a paired ground truth (which makes it more appropriate for a real-world application). The model combines contrast aware feature learning with perceptual reconstruction constraints; therefore, when using the proposed framework, less noise is amplified at the edges where there is no other way to maintain edge sharpness and texture fidelity. The encoder uses multi-scale illumination representation extraction methods; and then using progressively adding back the brightness and local contrast through feature fusing techniques the output of the decoder produces brightened image(s) with local contrast restored. A contrastive objective guides the encoder to produce discriminative illumination-aware features while the decoder reconstructs enhanced images with skip connections preserving fine detail. We present a comprehensive methodology, layer-by-layer architecture, full loss formulations with gradients intuition, training protocol and evaluation guidelines (BRISQUE, PSNR, SSIM and user-study design)