Abstract: Existing image analysis-based methods for exogenous mine fire detection are affected by complex mining environments and interference sources. Single-modal methods tend to misjudge light sources as fire sources, while multi-modal methods fail to utilize temperature information for fire source identification. Additionally, both methods have low identification accuracy under dust conditions. To address the above issues, an early fire source identification and anti-interference method for mines based on dual-spectrum imaging technology was proposed. First, the YOLOv10 model was used for real-time fire source detection on visible light images, and infrared thermal imaging was employed to obtain temperature distribution data. Then, Canny edge detection and image binarization preprocessing were applied to eliminate imaging differences between visible light and infrared images. Finally, the pHash algorithm was used to calculate the Hamming distance of the edge hash values between visible light and infrared images, and a threshold (Hamming distance≤25) was set to determine whether they represented the same fire source, thus effectively distinguishing fire sources from interference sources. The experimental results showed that under conditions without dust or interference sources, the accuracy of the early fire source detection and anti-interference method based on dual-spectrum imaging technology reached 98%, with a recall rate of 94%, outperforming the single-modal YOLOv10 (accuracy 97%, recall rate 86%). Under dust interference conditions, when 33% of the camera surface was covered by dust, the accuracy and recall rates were 85% and 80%, respectively. When 66% of the camera surface was covered by dust, the accuracy the recall rate were 70% and 65%, which were superior to both single-modal and multi-modal methods.