Real-Time Damage Detection Method for Conveyor Belts Based on Improved YoloX

To tackle the issues of high training expense, inadequate real-time functionality, and insufficient reliability in deep neural network-based visual detection of conveyor belt damage, which significantly limit the practical use of the detection model in industrial settings, this study proposes a novel YoloX-ECA damage detection method, based on the YoloX. Initially, the backbone network CSPDarknet is used for feature extraction, followed by invoking channel attention mechanism ECA on the enhance feature extraction network, which fuses multi-scale damage features and further amplifies the attention on damage target feature, facilitating the extraction of superior features to reinforce resistance to environmental interference. Additionally, the training strategy combines cross-domain transfer and intra-domain transfer learning to improve the training performance and robustness of the model. The experimental results demonstrate that YoloX-ECA has a detection accuracy of 95.65% and a speed of 30.50 fps. Moreover, compared to existing methods, the detection performance of the proposed method is effectively upgraded, resulting in a balanced accuracy and rate of conveyor belt damage detection.