具有散播型態之細胞蔓延於互動式前景擷取

Abstract

在現今熱門的圖形識別中，影像切割是一個必要的前處理過程。影像切割中的前景擷取根據相似性使同異質結合或分離，擷取出影像的特徵用於識別圖形。然而許多前景擷取無法從不同影像擷取出符合使用者需求的資訊，由於細胞蔓延法可以給予使用者提示，使擷取結果能夠符合使用者需求。細胞蔓延法無須分析和建立影像模型，模仿細胞強吃弱的生長蔓延性質，讓力量最強的提示往外蔓延，當蔓延到相異像素時降低力量並繼續蔓延直到被更強的相異像素阻擋為止。但是相似像素間些微差異，使像素蔓延到連通且遠距的相似像素時，力量略減過多無法阻止相異像素的蔓延；還有相似像素之間不連通無法蔓延過去。使用者需要給予更多提示在連通且距離遠的相似像素，和不連通的區域阻止錯誤蔓延，但會造成使用上不便。本論文利用蔓延力門檻限制像素往相異像素蔓延，並恢復已蔓延的蔓延力繼續往遠處的相似像素推進；接著利用蒲公英種子飛行的特性，散播些許前後景到不連通的相似區域，並落地生根長滿不連通的區域。實驗結果顯示本論文提出的方法可提高少許初始的切割準確性。

In today's popular pattern recognition, image segmentation is a necessary pre-treatment process.~Foreground extraction in image segmentation according to the similarity of features that combine homogeneous properties or isolated heterogeneous properties, extract the features in the image used to identify the photo.~However, many segmentation methods can not extract what users need from a different image objects, since users can be given prompt to GrowCut, so that the results can conform with user needs.~GrowCut do not need analysis of the nature of the image and to establish the image model, mimic cell the strong eat the weak by growth spread nature, let strongest hint spread out, reducing strength when spread to dissimilar pixels and continue to spread until it is far more distinct pixel barrier.~But excessive force was unable to prevent the spread of slightly reduced dissimilar pixels when the slight differences between similar pixels, so that spread to similar pixels and remote connectivity; and there is no communication between the similar pixels can not spread past.~Users need to inconvenience given more tips in similar pixels that connected also far distance, and also on disconnected similar pixels.~We use a threshold to limit spread to dissimilar pixel pixel and recovery force to spreading to the far and similar pixels;Then take advantage of dandelion seeds flying characteristics, spread little foreground and background to similar regions not connected, and covered not connected region.The experiment results presented in this thesis show that improve method can increase accuracy of a few initial seeds.