Multi-scale target intelligent detection method for coal, foreign object and early damage of conveyor belt surface under low illumination and dust fog

In the process of coal mining, foreign object such as gangue will inevitably be mixed into the coal, and ironware is easy to cause damage to the belt. The intelligent transportation system of coal mine needs to carry out integrated visual detection of foreign object such as gangue and belt damage. Aiming at the problems of unclear monitoring images and small target detection of early damage of belt surface caused by low illumination, uneven illumination and dust fog, a multi-scale target intelligent detection method for coal, foreign object and early damage of belt surface in low illumination and dust fog environment is proposed. Firstly, the low illumination and dust fog images are preprocessed by the contrast limited adaptive histogram equalization to enhance the image contrast. Then, by adding a shallow detection layer to highlight the detailed information such as the position and shape of the small target of the early belt surface damage, the performance of the early belt surface damage detection is improved, in the meantime some detection layers and corresponding feature extraction modules are removed to reduce the model without affecting the detection accuracy. Then, aiming at the problem of insufficient feature extraction ability of backbone network, a new P_Res2Block multi-scale feature representation module is constructed by using Partial Conv and Res2Net, and it is used to replace the Bottleneck of C3 module in backbone network to obtain a new P_RC3 lightweight multi-scale feature extraction module, so as to increase the receptive field of the model and enhance the attention to small targets. Finally, a lightweight and efficient channel attention mechanism is introduced into the backbone network and feature fusion network to improve channel utilization and early damage detection accuracy. The experimental results show that on the self-constructed low-illumination dust-fog image dataset, compared with the original YOLOv5s, the early damage detection accuracy AP_0.5 of the proposed method is improved by 10.00%, and the final accuracy is 91.30%. The number of model parameters, FLOPs and model size are reduced by 34.76%, 6.33% and 29.86% respectively, which proves that the improved method can solve the problem of small target detection in low illumination and dust fog environment while reducing the complexity of the model.