Abstract: Under high-speed conditions, coal gangue images suffer from degradation issues such as insufficient illumination, blurred details, and color distortion due to exposure limitations. This paper proposes a high-speed underexposed coal gangue image enhancement network based on a multi-task dual-branch architecture, termed HVI-MTDB-Net (HVI-based Multi-Task Dual-Branch Network for Underexposed Image Enhancement). The network first converts the original underexposed coal gangue RGB images to the HVI color space. Exploiting the light-color independence property of the HVI space, a multi-task dual-branch architecture is designed. This architecture utilizes a shared encoder to extract multi-scale features, integrating a Recursive Context Aggregator (RCA) within the encoder to enhance feature representation accuracy. The network achieves collaborative guidance interaction between the intensity branch (I-Net) and the color recovery branch (HV-Net). The I-Net introduces an illumination guidance module (IPG) to prioritize enhancement in dark regions while protecting edges, and the HV-Net incorporates a color fusion module (CFM) to simultaneously improve detail preservation and natural color restoration.Experimental results on a high-speed underexposed image dataset constructed in a real coal gangue collection environment demonstrate that HVI-MTDB-Net outperforms traditional color spaces such as YUV, Lab, and HSV, as well as mainstream methods like RetinexNet and EnlightenGAN, in image quality metrics including PSNR, SSIM, EN, and GM. In coal gangue recognition tasks, images enhanced by HVI-MTDB-Net significantly improve the accuracy, recall, and mAP of the YOLOv11n model. Notably, the mAP@0.5:0.95 increases by 6.6% compared to the next best method, validating the effectiveness and industrial application potential of the proposed method in underexposed visual enhancement scenarios.