Key Technologies Research On Computer Vision Inspection Of Mechanical Parts Based On Support Vector Machine

Computer vision inspection technology is computer vision based non-intrusive measuring technique integrated with the modern optical technology, electronic technique, mechanical manufacturing, control theory and computer technology. Visual inspection system can be roughly sub-divided into two categories: the system hardware design and the software system design.The system hardware design mainly includes lighting system design, camera system design, information channel construction, image acquisition and computer system selection; the software system mainly deals with the image processing technique, camera calibration technique.In this dissertation, with a view to improving the detecting precision and speed of visual inspection system, the key technologies of the image denoising, edge detection, sub-pixel detection and image compression, are conducted intensively and systematically.The support vector machine (SVM) is a kind of machine learning approaches in statistical learning theory fitting especially for small sample training, in which the generalization capability is improved by conducting the structural risk minimization (SRM). In dealing with the pattern recognition with small sample, nonlinearity and higher dimensional space, SVM gives many specific advantages which can be employed in other machine learning problems such as function estimating. In recent years, great importance was attached to the application of SVM in computer vision, further research on that basis to well exert the role of SVM in computer vision will be of vital importance for constructing computer vision detection system and elevating the measurement accuracy.The basic operator in image processing is the filter. A functional filter plays a very vital role to image processing. Conclusions are arrived at on an in-depth study of SVM principle, that the processing of signals regression process at evenly spaced points in discrete-time series utilizing least squares support vector regression is equivalent to a FIR filter, in which the constructing method of convolution mask is presented; and it is suggested that the FIR filter is a linear filter families differentiated by kernel function and penalty factor, and the derivative operator of every orders in strictly compliance with mathematical definitions is deduced. The using of such a filter can overcome, on one hand, the problem of wider margin of variation caused by utilizing difference method for approximate calculation because of not obtaining derivation operator when a filter is derived, on the other hand, the slower speed caused by matrix calculation in processing data using LS-SVM in conducting engineering problems.Images are usually corrupted by random noise in the process of camera imaging, digital image acquisition and image transmission. The random noises make it hard for the pixels value of images to reflect the optical properties of the object, which have led to a sharp drop in accuracy of measurement. Therefore corresponding measures have to be taken to decrease the influence from the random noise to measuring results. In this dissertation, theoretical analysis and experimental study of noise reduction by LS-SVR filter is conducted to reach the conclusion that the LS-SVR filter with RBF kernel function is equivalent to the gaussian filter in noise reduction effect, the parameters setting rule for LS-SVR filter are suggested also. LS-SVR based multiple-way switch universal image denoising algorithm is proposed, and the LS-SVR operator mask design technique and operating steps of the algorithm for 3-way switch and 17-way switch algorithm are included. Experiment shows that the proposed algorithms perform better results in the image denoising, the preserving for edges including its details and the computing speed, which provide with better image data recovery method for accurate measurement of vision checking system under strong impulse noise interference.Edges, as the general characteristics of an object, are is the base reference for geometrical feature extraction and analysis in vision inspection. The edge features may be changed with the random noise interference, an excellent edge detection method should possess both anti-noise ability and perfect preserving edge nature. Edge features analyzing combined with well-conditioned method for ill-conditioned image numerical derivative are conducted in this dissertation, the conclusion has been reached that the basic questions for edge detection are the solving for the derivative operators of smoothing filter. By analyzing canny's 1-d edge detection discrete criterion presented by Demigny, the expression of strict 1-d"good localization"criterion is proposed, on which, the grads based discrete expressions of criterion for the"good detection"and"good localization"of 2-d edge detection operator are created. The 2-d optimal operator corresponding to that criterion is solved in which the evaluation methodology in theory for 2-d edge detection operator is proposed that develops the edge detection theory by Canny. The approximate solution of 2-d optimal edge detection operator is obtained by LS-SVR, results of experiments and theoretical computation indicate that the obtained optimal operator is provided with both strong anti-noise ability and high precision of location. The experimental determination of optimum parameters for 2-d Canny optimal edge detection operator is conducted; Comparison between the results of 2-d Canny optimal edge detection operator and that of the 1-d Canny optimal edge detection operator shows that 2-d Canny optimal edge detection operator generated by 2-d Canny optimal edge detection operator using tensor product is not the optimal edge detection operator. By analyzing multi-scale in wavelet theory, a practical method realizing multi-scale detection by single operator is proposed.Usually the image pixels are roughly descripted by edge geometric characteristics, with the continuously increasing of the demand for the precision of vision inspection, the pixel accuracy will not satisfy the actual measuring requirement, employing sub-pixel method to express the edges of a body will remarkablely improve the detection precision of the part geometrical parameters. By analyzing the characteristics of sub-pixel edge detection, the sub-pixel edge detection is divided, in view of application, into image interpolation method and edge fitting method; considering the deficiency existed in optimizing function structure in the fitting method, the sub-pixel edge representation is presented by SVM in edge fitting; taking the sub-pixel edge parameters detection of the taper thread and the chain board for example, the input training sample set construction of SVM and sub-pixel edges description methods for the lines and circles are conducted in details.Image compression is one of the key technologies in vision inspection. Wavelet transform is an effective tool for analyzing non-stationary signals and also a major technique in image compression. The wavelet structure can be implemented using the wavelet lifting scheme, the reconfiguration can be directly obtained. By analyzing the wavelet lifting scheme and the Neville filter of the prediction operator for lifting scheme, the implementation of LS-SVR based Neville filter is proposed. The detailed parameters of LS-SVR for usual Neville filter are set. In view of the integer wavelet transform for a class of images, the typical LS-SVR filter is designed based on minimizing the wavelet energy, and the selecting principle for lifting wavelet prediction operator is analyzed also so.Research findings in this dissertation are theoretical significance and practical value to developing the image processing theory and promoting the application of computer vision in industry quality inspection.