Camshaft’s Journal Defects Detection and Classification Using Faster R-CNN Approach
Abstract
The camshaft is an essential part of a machine widely manufactured from wrought steel and nodular cast iron. One of the critical parts is a journal because this section should not have a physical defect on its surface. Several physical defects in the camshaft’s journal are identified in a typical gasoline machine camshaft. This work explores different state-of-the-art object detection methods and their applicability for camshaft’s journal and sprocket detection and classification tasks. Specifically, we implemented Faster R-CNN as part of Detectron2 using its base models and configurations. The results demonstrate that the X101-FPN base model for Faster R-CNN with the default configurations of Detectron2 is efficient and general enough to be applied to defect detection. This approach results in average correct detection score of 90.4% for test sets of the challenge. Though the visualizations show good prediction result, there are still some wrong defect detections. As a result, we compare the prediction results to the existing dataset and find some discrepancies.