使用於模糊前景影像與細胞影像之進階影像切割

Abstract

影像切割在電腦視覺和影像處理的領域中，扮演著相當重要的基礎處理角色，也有許多的相關應用如物件追蹤與影像壓縮等。現今已經有許多不同類型的影像切割演算法被提出，我們所提出的演算法是基於超像素(superpixel)、色彩、邊緣、質地、顯著值等特徵，進行分階段的超像素生長及適應性合併，用以將原始影像切割為使用者所希望的區域數。使用超像素生成局部區域資訊可以提升演算法整體運算效率，個別超像素的色彩及質地資訊是我們做為初步超像素生長的依據，超像素相接區域的邊緣梯度資訊則做為抑制超像素生長的因素，透過衡量原始影像的景物區隔性和計算顯著值，可以避免超像素被過度合併破壞物體資訊。在最後的適應性合併階段中，達到使用者輸入的原始影像切割區域數之前，合併過程會適應剩餘區域數。 若衡量原始影像的景物區隔性質方法應用在醫學影像上，可以估測出不同的細胞影像拍攝亮度，以調整套用在細胞切割演算法的亮度門檻值，可以改善經由影像亮度門檻值選取細胞區域的結果表現。

As a basic preprocessing procedure, image segmentation plays an important role in computer vision and image processing. There are many applications for image segmentation, such as object recognition and image compression. Recently, different kinds of image segmentation algorithms have been proposed. In this thesis, we propose an image segmentation algorithm based on superpixel, color, edge, texture and saliency information. The algorithm is designed to segment image into a certain number of regions assigned by the user. By using the superpixel information, one can improve the computation efficiency. The color and texture information of superpixels are mainly used in the superpixel growing process. On the contrast, the edge information on the boundary of two adjacent superpixels is used for determining whether the two superpixels should be prevented from merging. Saliency information is also a factor to suppress the merging process in order to keep the object integrity. In addition, we adjust the weight of edge, texture, and saliency information by measuring the foreground significance. In the adaptive region merging process, the merging criterion will be adaptive to the current region number. When the foreground significance is applied to medical cell image, we can estimate the imaging characteristic such that a better threshold can be chosen and further improve the cell image segmentation and tracing result.