Robust Active Contour Image Segmentation Models And Their Applications

Industrial computed tomography(CT) systems get projection data by scanning an object, and then reconstruct the tomographic images for image processing and analysis. Extracting objects of interest accurately and efficiently from these images are extremely important for the subsequent processing(such as target detection, dimensional measurement and reverse manufacturing, etc.).Magnetic resonance(MR) imaging is an effective non-invasive detective technology. It has been widely applied in the field of medical diagnosis as the human body can be free from radiation. MR imaging systems can get clear images of organs with more details. In addition, MR imaging can not only show the lesions of human organs, but also determine the function of each organ(such as brain and heart, etc.) accurately. Therefore, the accurate segmentation of various organs is very important for medical diagnosis.Image segmentation is the basic problem in image analysis and computer vision. Its purpose is to divide an image into disjoint regions, and then extract the target regions of interest with certain characteristics. As a key step from image processing to image analysis and understanding, the development of image segmentation technology is closely related to mathematics, computer science, pattern recognition, and other relevant disciplines and fields.In recent years, various methods have been proposed to solve practical segmentation problems. The active contour model(ACM), which based on curve evolution theory, has been of great attention. The ACM has been recognized and widely applied for its good performance and perfect theoretical basis, and successfully applied in industrial non-destructive testing, medical image analysis and other fields.Based on analyzing the existing region based active contour model, this thesis addresses the problems raised by the application of industrial CT and MR images. By combing some practical applications, we attempt to improve the current active contour models from several aspects, e.g., reducing the sensitivity to the initialization, enhancing the robustness to noise and improving the accuracy of the segmentation results.The main contributions and innovations of this thesis are summarized as follows:① An active contour model combing local robust statistics and correntropy-based K-means clustering is presented.In real world, images(e.g., CT images and MR images, etc.) with intensity inhomogeneities and noises are frequently due to the technical limitations of imaging systems and data acquisition conditions, so that the traditional image segmentation methods can’t get the accurate segmentation results of the targets. In this thesis, a model combing local robustness statistical and correntropy-based K-means clustering is proposed to improve the accuracy of the segmentation results for images polluted by intensity inhomogeneity and noise. The model consists of a global energy fitting functional and a local fitting energy functional. The global energy fitting functional is based on correntropy-based K-means method, so that the corresponding point in the cluster centers can be adaptive(image gray scale) sampled. This energy functional can emphatically sample these points whose intensity is close to the cluster center, improve the robustness to noise to some extent, and make the model insensitive to the initialization of level set function; the local fitting energy functional considers the local robustness statistics of the input image, which can effectively reduce the effect of noise to segmentation results. The use of local average intensity can make the model segment image with intensity inhomogeneities accurately. This thesis first gives the two-phase and multi-phase energy functional for image segmentation, and then gives the corresponding level set representation formulations. Experimental results of the synthetic images and real images demonstrate that our model is not only insensitive to the location of initialize contours, but also robust to different types and levels of noise.② A stable active contour model based on local area information is presented.For the images polluted by noise, an energy functional model based on Li-Kim model is presented by considering the mean value of local region intensity in this thesis. By using the differences of the intensity mean values between global image and local regions, we construct a robust image segmentation model based on local statistics for noisy images segmentation. Our model shows good robustness to noise by the introduction of local statistical information. The model can obtain stable minimum values. That is, if the initial level set function is bounded, then the final level set function is bounded. To get the minimum values, the evolution equation for the level set function of our model can be solved by semi-implicit difference schemes and analytical methods. The experiments of synthetic images and industrial CT images show that our model performs well for the images polluted by some noises(Gauss noise, impulse noise, speckle noise, etc.), and it is robust to different initialization schemes of level set function. The model can get more accurate of segmentation results, and the computation time can be greatly reduced.③ An improved active contour model for image segmentation and bias field estimation is presented.For real images, the intensity inhomogeneities can often cause many difficulties in image segmentation and image understanding(such as MR images). Bias field correction for MR images is a very important method before the diagnosis. In this thesis, an improved model based on Li’s model is presented for image segmentation and bias field correction. In our model, we consider the local regional differences between the real images and the estimated measurement images, and define cluster energy functional. Our model can get more accurate segmentation results and bias field estimation by introducing the local regional difference, and performs well for some low-contrast images. By utilizing the level set function regularization term, the level set function can be maintained the nature of the signed distance function in the evolution of curves, and the re-initialization can be eliminated. This thesis gives the two-phase and multi-phase clustering energy functionals and their corresponding level set function representation formulations. Experimental results of synthetic images and MR images demonstrate that our model can get more accurate segmentation results and more reasonable estimation of the bias field, and also can get better segmentation results for some low-contrast MR images.