Research On Automatic Detection Methods Of Welding Defects Based On Image Processing

Welding defect detection is an important way to ensure welding product quality. The traditionaldefect detection methods mainly rely on manual judgment of welding X-ray images with lowefficiency and high mistake rate. In recent years, the use of image processing techniques for automaticdetection of welding defects has been paid more attention by researchers. On the basis of previousresearch results, researches on welding defect detection involving image noise suppression, imageenhancement, image segmentation, feature extraction and classification methods have been done inthis thesis, and are described as follows:Firstly, a noise suppression method of welding defect image based on wavelet-contourlet transformand anisotropic diffusion is proposed. Wavelet-Contourlet transform is used to decompose weldingdefect image. The obtained low-frequency and high-frequency components are processed byimproved total variation (TV) and Catte_PM/kernel anisotropic diffusion (KAD), respectively. Finally,noise suppressed image is obtained by inverse wavelet-contourlet transform. A comparison is madewith three noise suppression method based on CTND, WSTV and CPMTV, the results show that theproposed method has better noise suppression performance and preserves minutiae more completely.Then, a welding defect image enhancement method based on non-subsampled shearlet transform(NSST) and nonlinear gain function is studied. A low-frequency sub-band and several high-frequencysub-bands are produced after decomposition of welding defect image through NSST, thenlow-frequency sub-band coefficients are adjusted by nonlinear gain function and a imagesegmentation method, the threshold and gain function are adapted by high-frequency coefficientswhich will be processed by the adapted threshold and gain function. Finally, welding defect imageenhancement is realized by inverse NSST. The experimental results show that compared with theenhancement method in contourlet domain, the enhancement method in non-subsampled shearlettransform (NSCT) domain, the studied method can improve the contrast of enhanced image, and theedges and detail of image are clearer.And then, a welding image segmentation method based on exponential cross entropy and improvedpulse coupled neural network (PCNN) is discussed. The area of weld is extracted by gray projectionalgorithm. Then, link weighted matrix and dynamic threshold function of PCNN are improved, andthe segmented image is obtained by improved PCNN, the optimal iteration times is determined byexponential cross entropy between images before and after segmentation. The experiments show thatthe segmentation results of defects are better than those of TECE and PCNN. Next, a welding defect extraction method based on improved CV model and PCNN in NSSTdomain is proposed. Firstly, a welding defect image is decomposed by NSST, the main region ofdefect is obtained through processing low-frequency component by using PCNN. Then,high-frequency feature image is constructed through low-frequency after background suppression andhigh-frequency, and improved CV model is used to search optimal contour of defect after coarsesegmentation. Finally, the final defect is extracted by fusing main region and fine contour of weldingdefect. The experimental results are given, compared with methods based on STCE, PCNN and NSCTcombining with PCNN, the extracted welding defect using the proposed method has more completestructure and optimal contour.Finally, a method of feature extraction for welding defect image based on contourlet transform andkernel principal component analysis (KPCA) by chaotic particle swarm optimization (CPSO) isproposed. Firstly low-frequency components and high-frequency components in a certain direction ofimages are extracted by Contourlet transform. Then, features of training samples and testing samplesof welding defects are extracted using KPCA after CPSO, respectively. Finally, the type of weldingdefect testing samples is determined according to the Euclidean distance between features of trainingsamples and features of testing samples. The experimental results that, compared with KPCA, waveletand KPCA, the proposed method can extract features more completely and has higher recognition rateand operating speed.