Novel roughness measurement for grinding surfaces using simulated data by transfer kernel learning

Abstract In conventional visual roughness measurement methods, constructing a relationship between an image feature index and surface roughness requires a large number of samples with a wide range of known roughness at uniform intervals as input for training or fitting. Considering these challenges, this paper has proposed a simulated data and transfer kernel learning-based visual roughness measurement method. In the proposed method, a virtual sample with specified roughness is first created via non-Gaussian surface digital simulation and three-dimensional entity modeling technology. After that step, a surface image of the virtual and processed samples is generated through image simulation and actual imaging experiments. Next, the image feature index distribution discrepancy between the simulation and actual domains is adapted by transfer kernel learning. A regression model is trained based on the simulated samples with known roughness, and is later generalized to the actual domain via a cross-domain kernel matrix to predict the roughness of the processed samples. To transfer the similar red and green mixing effects between the actual and simulation domains, a relative mixing degree index and a mixing region area index are designed based on the color information. By comparing these two indexes with the image pixel color difference index and image sharpness index, the feasibility and effectiveness of the proposed method are validated. The experiment results show that the proposed method can achieves an accuracy of over 90% based on the simulated data and transfer kernel learning. The proposed method provides a new improvement strategy for visual roughness measurement.