Improved Mask R-CNN for Defect Detection in Aluminum Alloy Rims
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Car factories have recently switched from steel to aluminum rims because they are more fuel-efficient, lightweight, and safe. This has made production more complex because these rims are more delicate and glossier and therefore more likely to have defects. Most rim manufacturers still use manual inspection or pattern-matching as a testing approach, which is known to be inefficient. This approach is also less able to meet Industry 4.0 demands, especially for producing alloy rims. One approach is to apply the newer You Only Look Once (YOLO) series network to inspect defects; although it has a faster detection speed, its ability to locate and identify smaller defects is insufficient, and defects can only be labeled with a rectangular bounding box, which makes it difficult to accurately obtain the extent of the defects. The traditional Mask Region-based Convolutional Neural Network (Mask R-CNN) can be labeled according to the defect morphology, but the accuracy is still unsatisfactory. Therefore, this paper proposes an improved Mask R-CNN network to overcome the limitations of existing networks. For more detailed and accurate feature extraction, the backbone structure is replaced by the ConvNeXt v2 module. The cascade region-of-interest head is incorporated into the back-end head structure to enhance defect recognition and classification accuracy and to facilitate more precise labeling and location. The proposed pixel-level area proportion algorithm is utilized to calculate each defect's area ratio based on the defects' masks in different colors. The experimental results demonstrate that, compared to conventional Mask R-CNN networks, the proposed method enhances average precision by 9.7%, average recall by 10.7%, F1 score by 10.1%, and mean average precision (mAP) accuracy by 6.5%. The proposed method also surpasses state-of-the-art YOLO series networks in performance and enhances the F1 score by 3.2% and mAP accuracy by 3.8%. The system also demonstrates acceptable detection speed and is equipped with a rotational photography method and graphical user interface design to realize real-time defect detection of an aluminum rim production line.