Using the system depicted in 'A hierarchical attention detector for bearing surface defect detection', a bearing surface defect image dataset comprising 859 defective bearing images with a resolution of 1024 $\times$ 1024 was created. According to the type of defect, the images are divided into four categories: depression, scratch, notch, and oil strain. Experienced quality inspectors on site annotated and reviewed the images using the LabelMe dataset annotation tool. The LabelMe surrounds the defect using a rectangular box, reflecting the defect location and type, and the annotations are saved in standard COCO format. The standard COCO format allows for fair comparison among different detectors and facilitates the evaluation of the proposed detector. These experiments use five-fold cross-validation to ensure the accuracy of the results. For each fold dataset, 80% of the bearing images are used for training, and the remaining 20% are used for testing. The model parameters are optimized on the training set. The final result of all the models is the average of the five test results. Notably, a single bearing defect image may include multiple defects.
baidu link and verify code: qfaw