Non-Uniform Low-Light Stereo Image Enhancement Based on Transformer

Low-light stereo image enhancement aims to improve the 3D perception accuracy of stereo vision systems in low-light scenes. However, existing methods do not fully address non-uniform low-light scenarios. To address this, we propose a Transformer-based method for non-uniform low-light stereo image enhancement, which consists of two stages: lighting adjustment and detail restoration. First, to mitigate the impact of non-uniform lighting on detail restoration, we introduce a global nonlinear light adaptive module. This module utilizes a Transformer-based method to adaptively generate gamma correction parameters, achieving nonlinear light correction and balancing lighting distribution. In the detail restoration stage, we designed a dual-branch mutually constrained cross-view interaction module to enhance the feature interaction between the left and right views, improving feature fusion. The main architecture adopts an encoder-decoder structure, and we introduce a Transformer-based global channel attention module in the feature extraction module to optimize channel weights, significantly enhancing feature extraction. We construct a non-uniform low-light stereo image enhancement dataset and compare our method with four single-view methods and three dual-view enhancement methods. Quantitative results show that our method performs the comparison methods on key image quality metrics, demonstrating higher structural fidelity and better brightness balance. Qualitative evaluation further demonstrates that our enhanced images exhibit significant advantages in detail restoration and lighting balance, effectively preserving textures while reducing noise and artifacts. Experimental results validate the effectiveness of our method in enhancing the quality of non-uniform low-light stereo images, particularly in color restoration and detail enhancement. The code and dataset are publicly available at: https://github.com/noportraits/NuLiSNet.