Flange surface defect detection based on the MROC-YOLO model

Abstract Aiming to address the issues of low accuracy and high leakage rate in multi-scale defect detection on flange surfaces in industrial scenarios, this paper proposes the MROC-YOLO algorithm, which improves upon YOLOv8m. A heterogeneous backbone network, termed MRNet, is constructed by integrating MobileNetV3 and ResNet to strengthen multi-scale feature extraction. Traditional convolutions are replaced with the ODConv dynamic convolution module, improving target area sensing and reducing missed detections. Furthermore, the CSCF module, which combines the SCConv module with two C2f layers, is introduced to minimize redundant information interference and enhance the detection accuracy of small target defects, enabling more efficient and precise detection. To tackle the scarcity of open flange defect data, the FL-DET dataset is constructed, encompassing four defect types: pits, patches, pitted surfaces, and scratches. Experiments show that the average detection accuracy of the model reaches 87.5%, which is 11.0%, 4.0%, and 13.9% higher than that of the original YOLOv8m in terms of precision, recall, and [email protected] indexes, respectively. Additionally, the computation amount is reduced to 20.0 GFLOPs, with a detection speed of 53.7 frames per second. These results demonstrate that the MROC-YOLO algorithm can effectively improve detection accuracy, reduce the missed detection rate, and is more suitable for real-time industrial detection.