結合標準化切割與色彩空間分佈特徵於自動化醒目性前景選取

Abstract

影像切割在電腦視覺與影像處理中都是一個重要且具挑戰性的問題。前景選取是影像切割中的一種，前景選取的目的是將一張圖片分成前景及背景兩個區域，前景的定義是一張圖片中離觀察者最近的部分場景和最醒目的位置。這種技術可以運用到物件偵測和物件辨識當中。現今已經有不少關於前景選取的方法，但是這些方法通常需要使用者的一些互動才能得出結果，使得使用上不方便。不同於先前的方法，在本論文中我們提出一個自動化前景選取的方法，以改善使用上的便利性。此方法利用了標準化切割和色彩空間分佈特徵。色彩空間分佈特徵可以標出醒目性區域所在的位置。因為醒目性區域符合前景的定義，我們可以藉著色彩空間分佈特徵來找出前景可能所在的位置。標準化切割藉由計算像素間的相似性和設定切割數目將影像分割成數個區域，這些區域未必已將前景和背景分割開來，我們提出將標準化切割和色彩空間分佈特徵結合以改善此問題。另外，本文藉著色彩空間分佈特徵提出兩個能自動化地在標準化切割區域中找到並合併屬於前景區域的方法，分別為門檻值法和卡方距離法。實驗結果證明了本文提出的方法不僅不需要使用者介入且結果也相當不錯。

Image segmentation is an essential and challenging problem in computer vision and image processing. Foreground extraction is one of image segmentation that separates an image into two regions, which are foreground and background. The definition of the foreground is the portion of a scene nearest to the viewer and a prominent position within the image. It can be used in object detection and object recognition. Recently, a lot of methods have been proposed for foreground extraction. However, the procedures of those methods need some interactions of users. It makes those methods inconvenient for users. Unlike previous methods, in this thesis, we propose a novel method for automatic foreground extraction. It uses normalized cuts and color spatial distribution feature. Color spatial distribution feature can indicate the position of the salient regions. Because the salient regions satisfy the definition of the foreground, we can find the position of the foreground by color spatial distribution feature. Normalized cuts can partition an image into several regions by calculating the similarity between pixels and setting the number of cutting. These split regions haven’t always separated foreground from background. We combine normalized cuts and color spatial distribution feature to overcome the problem. This thesis propose two methods to automatically find and merge the split regions of foreground among all the split regions of normalized cuts by color spatial distribution feature. These two methods are called ”thresholding-method” and ”chi-squared-distancemethod”. The experiment results show that the proposed methods do not need user interaction and perform well.