A study of computer vision for ground surface roughness evaluation

In the evaluation of surface roughness by computer vision technique, the pattern of illumination is generally correlated with optical surface finish parameters from the images. So this paper carried out experiments to investigate the effects of various factors and completed the optimum design of capture condition. Then we captured abundant sample images under appropriate experimental condition and chose to extract features of surface roughness in the spatial frequency domain which should be less sensitive to noise than spatial domain features. Therefore, artificial neural network (ANN), which took frequency-domain roughness features as the input, was developed to determine surface roughness by selecting the back-propagation algorithm. The built ANNs using these critical sets of inputs showed low deviation from the training data, low deviation from the testing data and high sensibility to the inputs levels. And the high prediction accuracy of the developed ANNs was confirmed by the good agreement with the results from traditional stylus method. Hence the proposed roughness features and neural network were efficient and effective for automated assessment of surface roughness.