Abstract: Limited exposure time in high-belt-speed coal-gangue sorting causes underexposed images. To address this problem, a Multi-Task Dual-Branch Network for Underexposed Image Enhancement, HVI-MTDB-Net, was proposed based on the Horizontal/Vertical-Intensity (HVI) color space. Underexposed coal-gangue images were mapped from the standard Red-Green-Blue (sRGB) color space to the HVI color space to effectively decouple illumination and color. A multi-task dual-branch network was constructed, in which a shared encoder was used to extract multiscale features and a recurrent context aggregator was designed in the encoding stage to improve feature extraction capability. During decoding and enhancement, an Illumination Enhancement Network (I-Net) and a Horizontal-Vertical Color Restoration Network (HV-Net) were designed to perform illumination enhancement and color restoration, respectively. Based on the exposure-time constraints of high-belt-speed imaging and image feature analysis, speed-specific exposure strategies were determined, and 200 groups of static paired dataset and a real-time high-belt-speed underexposed image dataset containing 2 400 images were constructed for network training and generalization testing, respectively. Experimental results showed that HVI-MTDB-Net achieved PSNR, SSIM, EN, and GM values of 21.875, 0.842, 7.931, and 4.786, respectively, on the test set, outperforming 10 mainstream image-enhancement methods, including RetinexNet and EnlightenGAN. When images enhanced by HVI-MTDB-Net were used as input to YOLOv11s, mAP@0.5:0.95 increased from 0.061 under the original underexposed condition to 0.551, and the frame rate reached 156 frames/s, meeting the requirements for detection accuracy and real-time performance under high-belt-speed conditions.