Design And Implementation Of Algorithms Based On CUDA Industrial CT Image Segmentation

CT (Computed Tomography) technology covers the knowledge of electronics, nuclear physics, precision machinery and computer application technology. It is used widely in many fields, such as the aerospace, nuclear, medical&biology, military, machinery, new material, customs, electronics, archeology and so on. The technology of industrial CT is usually used to detect product quality or safety structure. By processing and analysis of the CT image, the inner information of the detected objects can be obtained, so that it is possible to determine type of the inner material, and whether the object is qualified. Further, the technology of image segmentation is the essential work for processing and analysis of image.With the resolution of industrial CT device increasing, the quantity of CT image is very huge, especially to the stereo image data. Some stereo image data have10243pixels. Many algorithm of image segmentation are not so efficiency and can’t be applied in real project when they are used on so huge data. Further, in the CT image of one object, the distributions of intensity are often inhomogeneous. The global segmentation of CT image is also a crucial problem. Therefore, people focus on finding a method which can work out on CT image segmentation and can be accelerated by parallel computing using GPU.Parallel computing with GPU is a method which depends on multi-ALU on GPU and suit intensity of computing, that use CUDA or other programming language of parallel to accelerate the algorithm of serial algorithm.In this thesis, we adopt a global labeling based on energy minimization with the Potts model to segment the CT stereo data, and then we can accelerate the algorithm by parallel computing and complete global segmentation by using it. In practice, With GPU the parallel algorithm takes more less time than CPU. The following is major work in this thesis:(1) We adopt a global labeling based on energy minimization with the Potts model for defining the image segmentation. And then, we modify it to complete the segmentation of CT stereo image, and verify the algorithm for industrial CT image.(2) With analyzing the energy minimization formula and algorithm procedure, we implement the parallel segmentation procedure and programming it on GPU based CUDA language.(3) To verify that our parallel algorithm is practicable in real project, we practice it on CT image and CT stereo data. The experiments show that the global labeling base on Potts model can represent image segmentation very well and it can be implemented by parallel computing with GPU. So in the aspect of time efficiency, it can be applied in real project.