RBD-Net: robust breakage detection algorithm for industrial leather

For the sake of better achieving the productivity of leather damage detection in industrial production, this paper proposes a Robust Breakage Detection Network (RBD-Net) model for leather breakage detection. The model is an optimized model of You Only Look Once (YOLO) v5 to identify and detect the degrees of damage in leather and the type of damage in leather without any damage from the image. Firstly, the backbone network is replaced by Cross Stage Partial Densely Connected Networks (CSP-DenseNet), which can better achieve the reuse of features and prevent the loss of excessive gradient flow information; secondly, Bi-directional Feature Pyramid Network (BiFPN) is added in the feature refinement stage, which can better balance the feature information at different scales; finally, the addition of the Decision Network allows for capturing not only local shapes, but also global shapes spanning a large area of the image to better identify leather breakage in the image. By conducting experiments and making comparisons, it concludes that the method performs better than existing detection models in both breakage detection, and the accuracy of cutting, etched surface, brand stigma, hole and bleaching of five types of leather is 83.9%, 80.4%, 82.4%, 95.5%, and 89.4%, respectively, which satisfies the requirement for the balance of accuracy and robustness in industrial production, and also provides some ideas for other breakage detection research.