Volume - 6 | Issue - 4 | december 2024
Published
17 December, 2024
Enhancing low-light images under uneven illumination remains a challenging problem in computer vision. This study proposes an enhanced version of the Zero-Reference Deep Curve Estimation (Zero-DCE) model, named MS-DCE (Multi-Scale Deep Curve Estimation). The proposed model incorporates comprehensive architectural modifications and refined loss functions to improve performance. Specifically, multi-scale convolution is introduced to capture contextual information at varying scales, depth-wise separable convolutions are employed to reduce model parameters and computational cost, and traditional up-sampling is replaced with PixelShuffle to improve image resolution. Additionally, the loss functions are refined to mitigate overexposure while preserving natural colour consistency, thereby enhancing visual quality, particularly in regions with uneven lighting. Experimental results on the Part 2 subset of the SICE dataset demonstrate substantial improvements in image quality, with a 2% increase in PSNR and a 4% improvement in perceptual quality. These modifications not only enhance low-light image recovery but also provide a more efficient solution for handling complex illumination conditions.
KeywordsImage enhancement Neural networks Low-light enhancement Uneven lightening