Detection Of Ground Surface Roughness Based On Computer Vision

Surface roughness is one of important indexes reflecting microscopic error in geometrical form. And also, it is a famous parameter which is used widely to represent characteristics of surface. As the industry of mechanical manufacturing and detection technology develop, many groups are making intensive studies of various key technologies actively to realize quick and nondestructive examination of surface roughness.This thesis compared several ordinary measuring methods of surface roughness synthetically. Because the texture of ground surface displays strong randomness and insufficient directivity of distribution, a new detection method of surface roughness was put forward, through combination of computer vision and neural network technique. Roughness information of a line in traditional measurements would be replaced with information of a region by this method. And the survey of surface roughness implied a degree of repeatability, with memory of images and data information.Firstly, this thesis analyzed image characteristics of ground surface under different light sources profoundly, and chose condensing LED line source, whose model was LN-60, to be optimal selection of ground surface assessment through experiments. Hence, hardware system of surface roughness detection was established by adopting body microscope, CCD camera, image acquisition card and LED light source. Secondly, assessing the effects of various filtering methods and image enhancement approaches quantitatively according to profile curves of grey variation, the thesis developed effective image pre-processing programs, and put forward an detection algorithm of surface defects applied before surface roughness measuring by utilizing image segmentation based on grey-level thresholding. Therefore, we could recognize mass centric positions of surface defects and calculate properties of them, such as area, and so on. Thirdly, this thesis transformed ground surface images into frequency domain adopting two-dimensional Discrete Fourier Transform in order to implement further texture characteristic analysis. Then we discovered that power spectral radius, average power spectrum and central power spectrum percentage had approximately monotone relationship with the value of roughness, so these characteristics were extracted as inputs of BP neural network model to accomplish measurement of surface roughness, and the average percentage of accurancy could reach 93.8%. Hence, traditional stylus method could be substituted with this model under cetain conditions. Finally, this thesis developed detection software of ground surface roughness, taking LabVIEW as system platform, Visual C++ and MATLAB Neural Network Toolbox as auxiliary tools.