Voting Strategy Based Turbine Blade Cone-beam CT Image Segmentation Algorithm

Cone-beam CT(Computed Tomography) is an advanced three-dimensional industrial non-destructive detection method,which has important application value in the detection of micro-holes,internal cavities and defects.The detection accuracy of cone-beam CT mainly depends on the cone-beam CT imaging quality and the accuracy of the image post-processing system.Cone-beam CT image segmentation is the key link in the image post-processing process.The image segmentation and contour extraction accuracy directly affect the final detection accuracy of the CT system.Turbine blade material has high density and complex structure,which leads to serious noise,artifact,gray unevenness and low contrast in its cone-beam CT image.It is difficult to meet the high-precision detection requirements of turbine blades by using the existing segmentation algorithm.In this paper,to increase the image segmentation accuracy of turbine blade cone-beam CT images,the local window adaptive calculation,single-method voting image segmentation,multi-method voting weight optimization,high-precision three-dimensional image segmentation and point cloud extraction algorithms are researched deeply and systematicly.The main research contents and innovation contributions are as follows:(1)Adaptive local window image segmentation.Firstly,an adaptive local window image segmentation algorithm is proposed based on the characteristics of cone-beam CT slice image of turbine blade.Then,the necessary and sufficient conditions for determining the local window size of image segmentation are given.Based on the three criteria,i.e.the number of edge points,the difference between the target and the background and the peak numbers of local histogram in a local window,the local window size is calculated adaptively.Finally,the effectiveness and accuracy of the local segmentation algorithm are verified by simulated CT slice images.The results show that compared with other non-adaptive local window segmentation algorithms,the proposed method has the strongest segmentation ability for low-contrast and small-size defects,and the segmentation result has the highest contour accuracy,and the average edge precision reaches 0.9907.(2)Image segmentation based on single method voting.Aiming at the problem of insufficient correct segmentation probability in local window image segmentation algorithm,a new single-method voting based image segmentation algorithm is established by introducing voting strategy.The algorithm firstly guarantees the same number of votes for each pixel through image extension,and then votes all the pixels in the extended image in each local window containing the pixel.And uses the voting result instead of the original image to segment to improve the reliability of segmentation result.By introducing the CT image adaptability index of usual segmentation algorithms,the available segmentation algorithm set of CT image is screend out.Finally,taking the CT slice image of the ring workpiece as an example,the reliability and accuracy of the voting based single method segmentation algorithm are verified by comparing the accuracy of the extracted contour.The results show that the introduction of voting strategy can significantly improve the reliability and accuracy of segmentation results.Compared with other segmentation algorithms,the segmentation algorithm has higher contour accuracy and the average edge precision reaches 0.9891.(3)Multi-method voting weight optimization.From the perspective of maximizing the correct segmentation probability,the adaptability of different segmentation algorithms to CT images is comprehensively used.The multi-method combination voting segmentation idea and its weight optimization algorithm are proposed.Firstly,the influence of segmentation algorithm set and weight coefficient on the combination result is analyzed.Then the weight optimization model of multi-method voting is established,and the optimal combination voting weight coefficient is obtained by genetic algorithm.Then the effectiveness and accuracy of the multi-method voting weight optimization algorithm are verified by the CT image of the complex profile workpiece and the real hollow turbine blade.The results show that compared with the non-voting segmentation algorithm,the multi-method combination voting has stronger segmentation ability of noise,artifact and low-contrast small structure,and the segmentation result has higher contour accuracy,and the average edge precision reaches 0.9884 and 0.9727 respectively.(4)3D CT image segmentation and contour extraction.Aiming at the problem of low processing speed and insufficient segmentation accuracy of 3D CT image segmentation,a fast and high-precision segmentation algorithm based on knowledge of 3D CT images is proposed.Firstly,the spatial continuity of the CT image of the turbine blade is analyzed,and the CT image is preprocessed according to this feature.Then,the preprocessed image is three-dimensionally segmented and the three-dimensional segmentation is accelerated.Finally,to solve the inconsistency of the contour point cloud,a high-precision sub-pixel point cloud extraction algorithm based on multi-resolution image transformation is proposed.The CT images of complex cavity workpiece and real hollow turbine blade are segmented in the experiment.The effectiveness and accuracy of 3D segmentation and sub-pixel point cloud extraction algorithm are verified.The results show that compared with other non-three-dimensional segmentation algorithms,the proposed method segments noise,artifacts and low-contrast small structure targets better,and acquires higher contour accuracy.The point cloud acquired by our proposed method is consistent.For the complex cavity workpiece,the edge precision reaches 0.9851.For the blade workpiece,the edge precision reaches 0.9843,which is 0.0116 higher than the pixel level result.