Research On Technology Of Industrial Computed Tomography Non-destructive Testing Based On Machine Vision

The rapid development of industrial computed tomography(CT) technology has provided a great convenience for non-destructive testing(NDT) of industrial equipment. The processing after imaging is very important for improving image quality and effective detection. Aiming at constructing a complete machine visual system to realize the automatic detection of industrial equipment on line, the involved technologies including image preprocessing, image segmentation, edge detection and so on, are researched. The main work is as follows:Firstly, a method of non-local means(NLM) denoising for industrial CT images based on kernel singular value decomposition(KSVD) and similar pixel labelling is proposed. The KSVD is applied to extract the main algebraic features of similar-window patches for sparse representation. Based on this, the similarity is measured and the calculation is reduced. The pair of similar pixels is labelled so as to avoid repetitive computation of similarity between two pixels. Thus, the size of search-window decreases by half compared to the normal NLM method. The experimental results demonstrate that the proposed method has good performance in 3 aspects, namely, subjective visual denoising effect, peak signal to noise ratio(PSNR) and processing speed. Besides, the proposed method has obvious advantages against the normal NLM method, the principal neighborhood dictionaries(PND) method and the Zernike moment based NLM method.Then, an image enhancement method in shearlet domain is discussed. An image is decomposed by non-subsampled shearlet transform(NSST) into high-frequency part and low-frequency part. The high-frequency part is processed through non-linear transform in order to enhance image edges and suppress noise. The low-frequency part is handled by local-block enhancement, in which the human eye perception information fidelity is introduced to solve the distortion problem of boundary pixels. Compared to the double plateaus histogram equalization method, fidelity constraint method, contourlet based fuzzy enhancement method and non-subsampled contourlet transform(NSCT) based adaptive threshold method, the proposed method obtains enhanced images with better visual effect and has a 50% improvement averagely in definition, local contrast and global contrast. The processing time is just 10% compared to the NSCT based adaptive threshold method which has a suboptimal performance in the quantitative indicators mentioned above.And then, an adaptive thresholding method for industrial CT images is proposed. Firstly, the reciprocal gray entropy is defined, and the one-dimensional reciprocal gray entropy thresholding formula is derived with the uniformity of within-cluster gray level considered. Thus the drawback of undefined value of Shannon entropy is avoided.. To improve the anti-noise performance, the method is extended and the two-dimensional reciprocal gray entropy formulae for threshold selection are derived. The decomposition algorithm of two-dimensional reciprocal gray entropy is proposed to reduce the calculation. Directly searching the two-dimensional threshold is replaced by searching two one-dimensional optimal thresholds so that the computational complexity is reduced in a great degree. Compared with the improved Otsu method, the two-dimensional maximum Shannon entropy method based on particle swarm optimization(PSO) and the two-dimensional reciprocal entropy method based on histogram oblique segmentation and niche chaotic mutation particle swarm optimization(NCPSO), the proposed method achieve better segmentation results for industrial CT images.Subsequently, an industrial CT image edge detection method based on NSST and guided filtering is researched. Firstly, the guided filtering is introduced to improve the Canny edge operator and a preliminary detection result which is regarded as the edge of low-frequency image is obtained. And then, the NSST is implemented for image decomposition and the high-frequency coefficients are extracted. The modulus maximum detection is performed for high-frequency coefficients to find more edges. The modulus maximum values are further adjusted to eliminate false edges depending on the property of coefficients of edge points under different decomposition conditions. At last, the high-frequency and low-frequency edges are fused to get the final edge image. The proposed method is compared with common Canny method, modulus maximum method in wavelet domain and modulus maximum method in NSCT domain. The experiments prove that the proposed method is able to detect edges more accurately and has the least difference with reference image, and thus has the higher FOM.Finally, a defect detection method for industrial CT images based on visual attention mechanism is proposed. The guided filtering is adopted in place of Gaussian filtering to generate scale space avoiding destruction of edges. Each feature map is viewed as a graph to construct a Markov chain and the equilibrium distribution of the Markov chain is taken as the saliency map. In view of the characteristic of industrial CT images, feature maps are not extracted from color channel. The simply mean-manner of saliency map fusion is replaced by proportion-manner according to the different properties of saliency maps. An adaptive thresholding method is implanted in the extracted saliency area to prevent the inaccuracy of manual thresholding. The comparison among the Itti method, the graph based visual saliency(GBVS) method and the proposed method shows that, the proposed method can obtain clearer and more accurate saliency map and has a better defect detection performance for industrial CT images.