Study on Visual Detection Method of Multi-scale Damage to Conveyor Belt Under Complex Background

Belt conveyors are one type of the most important transportation devices for continuous transportation of bulk materials in bulk cargo piers, whose operating environments are complicated. The belt damages change in a wide scale range and multiple scales coexist. Thus, the reliability of damage detection methods is challenging. A Yolov5-based improved EMA-YOLO belt multi-scale damage detection method is put forward here. First, a high-efficiency multi-scale attention module EMA is introduced in view of there being complex background interferences in belt damage detection samples so that important areas shall be more concerned in our model. The multi-scale features can be extracted at different levels and the resistance ability of background interferences may be enhanced. Additionally, the feature fusion network is improved. The weight of learning is introduced by means of a simple and efficient weighted bidirectional feature pyramid network (BiFPN) to learn the weights of different input feature layers and modular repeated application is performed so that our model can simply and quickly fuse multi-scale features. Thus, the training performance of our model can be improved. Our measurements indicate that the accuracy of our detection method is up to 98.51%, and its detection speed is up to 82.26 FPS. Compared with the existing methods, our method can be more reliable and real time.