Deep learning and image processing-based method for automatic estimation of metal-machined surface roughness grades

Abstract

Surface roughness is one of the critical technical requirements of precision machining engineering. Traditional assessment methods, such as standard sample comparisons or contact roughness measurement devices, have long shown limitations. The case of this study is in the context of machining workshops equipped with computer numerical control systems (CNC), where turning and milling methods account for an average of 60 % of the machining process. Based on convolutional neural networks and image processing techniques, this study proposes a method and a hardware structure to support non-contact roughness grade evaluation through surface texture images. The device is suitable for medium- and small-sized machine parts, meeting the practical production context of CNC machining workshops and the ISO 1302:1992 roughness grade classification standard. The training data were generated from images of surfaces with known roughness levels within the Ra 0.4–3.2 µm. The model achieved an average accuracy of 85.83 %, indicating the feasibility of applying convolutional neural networks and image processing to determine and assess the quality of machined surfaces.