Detection of coal gangue based on spectral technology and enhanced lightweight YOLOv7-tiny

Underground coal gangue sorting is a critical component of modern intelligent coal mining, holding significant importance for the preservation of natural resources and the ecological environment. Traditional methods of underground coal gangue sorting suffer from issues such as low efficiency, limited applicability, and substantial resource wastage. Addressing these challenges, this paper employs multispectral technology to gather spectral data of coal and gangue in various wavelengths. Based on the identification accuracy of coal gangue images in different wavelength bands and the correlation of spectral data, the optimal three wavelengths out of 25 are selected to construct a pseudo-RGB (Red, Green, Blue) image. Furthermore, building upon YOLOv7-tiny, an improved lightweight coal gangue recognition method is proposed. Experimental results demonstrate that the improved lightweight model has a computational load of 11.5 GFLOPs, merely 88.5% of the original model's load. The model's detection rate is 77 frames per second (fps), a 23 fps increase compared to the original model. Precision, recall, and average accuracy reach 98.7%, 97.1%, and 98.8% respectively, indicating a 1.5%, 0.2%, and 0.5% improvement over the original model. The approach effectively mitigates instances of omission, enhancing model accuracy and portability.