Study On Feature Extraction Of Industrial CT Image And Volume Data Based On Multiscale Geometric Analysis

As an advanced nondestructive testing (NDT) technology, industrial computed tomography (ICT), which refers to reconstructing cross sectional image (CT image) and volume data (CT volume data) of an object from projections taken at different directions around the object, is especially useful for analyzing complex and close objects. Feature extraction is of initial importance in ICT data processing. Among the varied features of ICT image and volume data, line (contour of the object) and surface (contour surface of the object) are two of the significant ones, the extraction efficiency of which will directly affect the performance of next data processing (such as measurement, reverse engineering, etc.).Following wavelet, multiscale geometric analysis (MGA), also referred to as"beyond wavelet", is one kind of useful tool for analyzing data, especially for higher dimensions data. These methods can generally be sorted into two groups: directional basic functions or frames on one hand, such as curvelets, contourlets, bandelets; adaptive geometry-based approaches on the other hand, such as wedgelets, beamlets, platelets. The research work of this dissertation belongs to the latter group.The research of this dissertation is supported by the National Science Foundation of China (No. 60672098, No. 60972104 etc.). This dissertation concentrates upon practical problems of ICT, using methods of MGA. It not only designs some fast computation methods of MGA, but also proposes solutions to line feature extraction and surface feature extraction of ICT image or volume data. The contributions and innovations of this dissertation include several parts as follows:①Designs of some fast computation methods of MGA. Each method of 2D beamlet transform, 3D beamlet transform, 2D wedgelet decomposition, and 3D wedgelet decomposition has its own enormous basis functions. If computing them one by one, there is high computation cost. After analyzing the component and relation of each approach at mono-scale, a fast computation method is proposed, which improves the computation efficiency.②For ICT noisy image of crack (with approximate linear or piecewise linear characteristic) detection, method based on beamlet is presented. Typical detection methods use"point bases"to analyze image, which have preprocessing, such as denoising etc. However, beamlet uses"line bases"to analyze image, which has robustness to noise. Furthermore, considering beamlet is just beeline, in order to enrich the line bases, beamlet and quadratic curve are introduced as new"line bases"for crack detection. Compared with the method of laplace, sobel, canny or wavelet, the proposed methods can extract the crack more efficiently.③For ICT noisy volume data of crack surface (with approximate plane or piecewise plane characteristic) detection, method based on finite plane integral transform (FPIT, it is presented by this dissertation.) and planelet is presented. Both FPIT and planelet use"plane bases"to analyze volume data, which have robustness to noise. FPIT computes individual voxel of planes integral, which is suitable for ascertaining approximate location of crack surface. However, planelet computes individual dyadic cube of planes integral, which is suitable for precise crack surface extraction. As a result, this dissertation uses 3×3×3 FPIT plane templates to find out approximate locations of crack surface first, and then uses planelet to extract it precisely from approximate locations. Compared with the method of 3D facet, C-V, and 3D wavelet, the proposed method can extract the crack surface more efficiently.④For ICT volume data of line feature extraction, method based on improved 3D finite line integral transform (I-3D FLIT) is presented. Based on the concept of FLIT, this dissertation generalizes its application to the case of 3D, attaining the formulas of 3D FLIT. Due to the incontinuity of many lines defined in 3D FLIT module, this dissertation improves it to continuous lines called I-3D FLIT module. On the basis of it, image fusion and morphological processing are combined to extract the line feature. Compared with the method of proposed I-FLIT or 3D wavelet, the method of proposed I-3D FLIT can extract finer and full line features of volume data.⑤For ICT volume data of line feature extraction, method based on 2D wedgelet is presented. Method based on I-3D FLIT need to analyze data voxel individually, while method based on 2D wedgelet uses"wedge bases"to analyze individual dyadic square of the data, which is suitable for extracting line feature in the form of different conjoint gray regions. Compared with the method of 3D wavelet, the proposed method can extract finer and full line features of volume data. Compared with the method of I-3D FLIT, the proposed method can describe the line features of volume data more compactly.⑥For ICT volume data of surface feature extraction, method based on 2D/3D wedgelet are presented. There are two strategies of surface feature extraction for ICT volume data. One is dividing volume data into slice groups at certain orientation first, and then extracting line feature on each slice, and combining line feature and attaining surface feature finally. The other one is extracting surface feature directly in volume data. Most of the surface features of ICT volume data exist in the form of different conjoint gray regions. If volume data are divided into slice groups, surface feature changes into line feature that takes the form of different conjoint gray regions. 2D/3D wedgelet uses"wedge bases"to analyze individual dyadic square/cube of the data, which is suitable for extracting surface feature in the above form. Compared with the method of 3D wavelet, the proposed methods can describe the surface features of volume data more compactly.