Dual-Branch Multiscale Optimization Network for Enhancing Low-Light Images in Rail Transit Obstacle Detection and Segmentation

The existing deep learning-based low-light image enhancement models have demonstrated their validity on many benchmark datasets. However, it is not easy to enhance the brightness, color, contrast, and other information of images while maintaining the quality of image details with these models. To enhance the visual perception of images under low-light conditions at night, improve driving visibility, and increase the accuracy of obstacle detection and segmentation, a dual-branch multiscale optimization (DBMO) network is proposed in this article. The global branch employs a multiscale network based on the transformer, which preserves high-resolution input images while effectively integrating multiscale features to capture global image information. Meanwhile, the detailed branch introduces an adjustment denoising network based on wavelet transform. It performs noise suppression and detail enhancement on the high- and low-frequency information obtained from wavelet decomposition, thereby enhancing image details while balancing brightness and contrast. Finally, the feature information extracted from both branches is adaptively weighted and fused to produce the final enhanced image. The experimental results on a real rail transit obstacle dataset demonstrate that DBMO significantly improves images’ overall brightness and color balance and achieves the highest accuracy improvement in rail transit obstacle detection and segmentation. Compared to baseline models YOLO-v8 and Deeplab-v3+, obstacle target detection and semantic segmentation accuracy improved by 3.6% and 7.65%, respectively. This model can be applied to nighttime train assistance systems and shows promising potential in obstacle detection and segmentation tasks.