Abstract
As the core components of permanent magnet motors, the surface defects of magnetic tiles can directly affect the working performance of the motors. There are various types of defects in magnetic tiles, such as chipping and wear. Traditional magnetic tile defect detection mainly relies on manual inspection and faces challenges like low detection accuracy and high cost. Moreover, most defects on magnetic tile surfaces are located on curved surfaces, leading to an uneven distribution of defects and tiny features, which makes accurate defect localization challenging. To solve these problems, this study proposes a novel magnetic tile defect detection algorithm called YOLO-RDM. First, we apply DOConv to the neck network. By using a lightweight convolution method, we replace the traditional convolution in the original network, thereby improving the feature extraction ability of the model and achieving lightweight processing. Second, we design an RPA Block to improve the C2f module. By introducing a parallel attention mechanism, we enhance the feature extraction ability of the algorithm. Finally, we replace the original backbone network of YOLOv8 with the MogaNet network. MogaNet is a module that aggregates contextual information, enhancing the network's discriminative power, learning efficiency, and ability to capture defect features in images. The experimental results show that the mean average precision (mAP@0.5) of the improved model reaches 95.0%, which is 4.8% higher than that of the original model, and its inference time is less than 5.6 ms. It also has obvious performance advantages compared with other object detection models. In addition, it achieves good recognition results on the NEU metal surface defect dataset. It can be proven that the YOLO-RDM model has strong recognition and generalization abilities and can be used in practical applications of magnetic tile defect detection.