Research On Vision Inspection Based On Wavelet Analysis

Vision inspection is an application of machine vision theory and technology in the fields of inspection. It has become one of the very fast developing fields in the instrumentation science. As we have known, image processing is a critical part in the vision inspection system. Due to its unique advantages, wavelet analysis has been widely used in image processing. The applications of wavelet analysis to image processing of vision inspection are studied in this thesis. Researches are mainly focused on the following aspects:(1) In this thesis, it was proved that the characters of Pulse noise and Possion noise are same as Gaussian noise in wavelet domain, which means the amplitude and variation of Pulse noise and Possion noise in wavelet domain decreases as the scale increases. It theoretically ensures that denoising method for Gaussian noise can also be used for Pulse noise and Possion noise.(2) In this thesis, the denoising method based on threshold-filtering of wavelet coefficients is introduced first, and then a denoising method based on threshold-filtering of wavelet packet coefficients is proposed. According to theory and simulation experiments, it was proved that the performance of the denoising method based on wavelet packet coefficients is better than based on wavelet coefficients. Furthermore, two critical problems, how to determinate the wavelet packet decomposition and how to choose the threshold in the denoising method based on wavelet packet coefficients are pointed out.(3) In Donoho's nonlinear threshold-filtering of wavelet coefficients method, its threshold is acquired on the hypothesis that the noise obeys the normal distributing, which does not agree with reality. In this thesis a denoising method based on self-learning threshold-filtering of wavelet coefficients or wavelet packet coefficients is proposed.(4) The capabilities of several edge-detecting-operators are compared with the theory of data smoothing and Canny's edge-detecting-criteria. So we can choice the edge detecting operator in wavelet multi-scale edge detection.(5) The effective scale ranges of multi-scale edge detection based on wavelet are analyzed. Since the variogram function can efficiently expresses structural and statistical property of image data, it can discriminate the noise and the edge of image.Based on the property of the variogram function, a scale-adaptive edge detection algorithm is proposed.(6) After studying and analyzing characteristics of current sub-pixel edge detection methods, an new sub-pixel edge detection algorithm based on wavelet coefficient expectation is proposed. Based on the theory, simulation and experiment, we have proved that this algorithm have high precision in edge location and good anti-noise ability. At the same time, the speed of the method is higher than the algorithm based on the moment.(7) In industrial vision inspection, many edges of inspecting objects are. some kinds of roof edges, but current sub-pixel edge detection methods are only applied to step edges. Because roof edge is different from step edge, their detection methods are different. So we proposed two sub-pixel edge detection algorithm for the roof edge detection. The first algorithm is expectation of the wavelet coefficient derivative and the second is zero-cross of the wavelet coefficient. Based on the theory, simulation and experiment, we have proved their capabilities of sub-pixel edge detection are effective.(8) A wavelet invariants, moment calculated by wavelet coefficient is deduced. Through the experiments the precision of the approximation is analyzed.(9) A wavelet moment invariants is deduced and a new algorithm of wavelet moment based on FFT is proposed. Experimental results prove that wavelet moment invariants are superior to Zernike's moment invariants, Hu's moment invariants and moments calculated by wavelet coefficient for character recognition.