Detection of surface roughness of mechanical drawings with deep learning

Engineering drawing inspection is important to CAD modeling of mechanical parts. Traditional inspection methods mainly rely on manual analysis by using the CAD software, which requires expert knowledge and massive time. In view of simplifying the analysis for non-experts and improving detection efficiency and accuracy, this study proposes a generic approach combining object detection and image recognition methods to identify surface roughness of mechanical drawings. For both the object detection and image recognition methods, deep learning models with different backbone networks are trained and tested independently. Experimental results show that a combination of Faster-RCNN with ResNet101 as backbone network, and SSD with ResNet50 as backbone network achieves the best performance under our evaluation metrics.