Application of visible and infrared image feature fusion based on YOLOv10n in coal and gangue recognition

To address the challenge of inaccurate coal and gangue recognition in complex environments, this study proposes a dual-channel input YOLOv10-based method for visible and infrared image feature fusion, achieving efficient recognition and detection of coal and gangue. Comparative analyses were conducted between the fused network model and the original YOLOv10n model trained and tested separately on visible and infrared images. The experimental results show that the fusion network is better than the single-modal group in Precision, Recall and mAP50 indexes, and it has obvious stability in various complex scenarios, which proves the effectiveness of the feature fusion strategy. In particular, the computational complexity and parameter quantity of the fusion model on GFLOPs are about half that of the baseline model, which is more suitable for deployment on edge devices. Furthermore, to further reduce the computational complexity, the Ghost module for lightweight design is introduced. Systematic ablation experiments were implemented through multiple comparative strategies, including replacement of attention mechanisms and loss functions. Ultimately, by substituting the PSA attention mechanism with the SE attention mechanism and adopting WIoU as the loss function, an optimized model was successfully obtained that maintains or improves recognition accuracy while achieving enhanced efficiency.