Defect detection plays an important role in implementing zero-defect manufacturing (ZDM) and improving the sustainability of manufacturing systems. The remarkable diversity of gear types, the inhomogeneity of end-face structure, as well as the small size and multi-scale of defects, are the common problems confronted during the metal gear end-face defect detection, which leads to poor performance of existing detection methods in terms of detection rate and accuracy. To address the problems above, this study proposes a cascaded combination method SR-ResNetYOLO to automatically detect the defects by region extraction and multi-scale fusion of sampled features under 16X. To obtain more effective features, this study proposes the visual-saliency-based method to extract the machined area image, eliminating the interference between the invalid features of non-machined areas and edge burrs and reducing thereby the image complexity. Subsequently, establish a 16X down-sampled feature extraction backbone network (ResNet-21), to efficiently obtain the high-resolution features of the defects by using the machined area images as input. With the multi-scale fusion module, the min-scale feature map, output by the ResNet-21, fuses at the medium- and large-scales. Finally, the three-fused-scale feature maps are classified and located by the location and classification module. The proposed method achieves satisfactory performance in terms of the mAP and recall rate, which are respectively 96.66% and 97.07%, and the average computation time of the detection for per image is 0.12 s, which can effectively detect small size and multiple scale defects of metal gear end-face.
Read full abstract