Abstract: The underground coal mining environment is complex and typically characterized by low illumination and high concentrations of fog and dust. Under such working conditions, manual inspection is difficult and inefficient, so machine vision–based detection methods are widely adopted. However, the machine vision detection of foreign object intrusion on belt conveyors under low-illumination conditions has long faced severe challenges. Visible light images lack sufficient information, while infrared images, rich in grayscale data, lack detailed texture, leading to low recognition accuracy and missed detections. To address the difficulties of visual perception in such low-light conditions, an image processing method based on multi-modal information fusion theory, which integrates infrared and visible light images, is proposed. The proposed framework consists of two core components. First, at the front-end, a Coal Pile Adaptive perception registration algorithm (SuperGlue-CPA) is employed. Utilizing an elliptical physical constraint weight mask and a layered homography matrix strategy, this algorithm reduces registration error by 33.51% and increases the effective feature ratio to 95.33, significantly improving registration accuracy. Second, at the back-end, a Multi-scale Scene-adaptive sparse semantic image Fusion algorithm (MSPFusion) is proposed. This algorithm enhances feature representation capability by introducing a Global Grouped Coordinate Attention module. It also improves the semantic injection method to achieve global feature retention and parameter decoupling. Furthermore, its semantic perception network incorporates multi-scale convolutions to address the scale variation of foreign objects. These improvements result in a 1.4% increase in mean Intersection over Union (mIoU) and a 9.4% increase in fused information quantity, markedly enhancing the overall capability of the fusion algorithm. Experiments demonstrate that the proposed method significantly improves the visual quality of fused images and substantially enhances the salience and recognition confidence of foreign objects. This work provides a new approach for the visual applications of underground inspection robots, thereby reducing belt conveyor accidents caused by foreign object intrusion in low-light environments. It holds important application value for ensuring coal mine safety and enabling intelligent transformation.