Wheel hub defect detection based on the DS-Cascade RCNN

Abstract

At present, object detection methods based on machine vision have been widely used in the field of industrial defect detection. Wheel hub defects are characterized by multiple scales and complex types. The location, size and affiliation of different defect marks are different, so it is difficult to establish an accurate wheel hub defect detection model. Therefore, a wheel nuclear hub defect detection method based on the DS-Cascade RCNN was proposed. To effectively locate multiscale s, a spatial attention mechanism was added as a wheel hub defect location enhancement module. Then deformable convolution is added, and the position and size of the convolution kernel are adjusted dynamically according to the shape of wheel defects. Finally, the pruning algorithm is used to optimize the improved model and compress the model space without losing the accuracy. The model was evaluated under the wheel dataset. Experimental results show that the proposed method can effectively detect six kinds of wheel hub defects, and the mean Average Precision (mAP) is 95.49%. Multiscale defect location and defect category estimation are realized, which meets the requirements of wheel hub detection in actual production.