關於影像輪廓的主動演進和離散型的切割法

Abstract

主動輪廓切割模型，暱稱蛇，是影像切割方法裡的一個很強的工具。 因為它可以針對感興趣的區域提供一條連接的輪廓。所以，較一般切割方 法而言，省去了需設法將找出的輪廓連接的工作。然而，此模型亦有它相 對的缺點，如邊角的切割，雜訊的影響等等。這些都是主動輪廓切割模型 難以克服的缺點。為了解決雜訊影像的輪廓切割問題，本研究提出兩種主 動輪廓切割模型，一是主動演進型切割法，二是主動離散型切割法。前者 主要是針對邊角的偵測，後者則是針對雜訊的影像結合了我們之前所發展 出的早期視覺模型的方法。 後者是離散化地在距離圖上尋找真正輪廓， 而不是一個像素一個像素地尋找。因此它可以克服雜訊的影響，以及加速 了收斂速度。除了主動演進型及離散型兩種方法之外，本研究還提出一套 新的適應性權數調整方法。原則是不斷地透過參數的調整功能，盡可能保 持各個能量函數的平衡。再運用整條搜索方向全部的影像訊息，提供輪廓 靠近真正邊際的一個推進力量。使用這些技術，我們可以切割出模擬的雜 訊影像的角形輪廓，並且可以切割臨床的超音波醫學影像。 Active contour model, also known by the nick name, snake, is a powerful tool for image segmentation. Since it provides continuous boundaries for regions of interest, a snake model has the advantage over edge-detecting approaches for segmentation in that no edge linking would be required.Powerful as it is, the snake model suffers several cons, e.g., corner- fitting problem, struck by noise pixels, etc., which limits itsusage, especially, on a noisy image. With the ultimate goal to accomplish segmentation on a noisy image, thisstudy provides two novel snake models to solve the corner-fitting and noiseproblems. One is evolutionary snake-balloon model and the other is discretesnake-balloon model. The evolutionary snake- balloon model and the other is thediscrete snake-balloon model. The evolutionary model is an effective approachto catching corner points while the snake is moving toward the desired boundaries. The discrete model, on the other hand, circumvents the noisy problems by incorporating our recently developed early vision model.Instead of deforming on the image of interest pixel-by-pixel, the snakesearches for actual boundaries on the distance map discretely, which not onlyminimizes the noise effect but also speeds up convergence significantly. In addition to evolutionary and discrete models, a new adaptive methodologyis proposed in this thesis to determine the weighting factor of each term in the snake energy functions such that the snake may cope with local minima more effectively. The weighting factorsare adjusted adaptively to balance all energy forces according to preset ratio among these forces to provide appropriate forces for deformation and stopping on the actual boundary. Experiments have been carried out to verify the proposed evolutionaryand discrete snake-balloon models. For the evolutionary snake-balloon model,various phantoms filled with gaussian random noises are used to simulate noisy images. By controlling the means and standard deviations of the gaussiandistributions, the noise-resisting capability of the evolutionary model for different signal- to-ratio has been examined. For the discrete snake-balloon odel, clinical ultrasound nimages are used to demonstrate its superiority over the conventional snake models. In most cases, the discrete snakes converge in less than ten steps which is substantially faster than the conventional ones.