基於圖像分割的自動化前景選取

Abstract

影像切割在電腦視覺中, 是一個歷史悠久且極為重要的領域。影像切割希望特徵同質者結合、異質者分離, 且能自動切割區域為有意義的形狀。大量文獻回顧顯示 GrabCut 是目前效能較佳的方法, 它是先定義能量函數再利用圖論演算法進行切割的前景選取方法。GrabCut 僅將影像分割為前景與背景, 且切割正確率高, 值得一提的是 GrabCut 採用區域項與邊界項定義能量函數, 使得此方法符合人類視覺。然而 GrabCut 僅為半自動切割法, 它須由使用者圈選背景種子點後才可進行切割, 使用上極為不便, 因此本論文希望以一個貼近邊緣的初始方框自動切割。 為達此目的, 必須先解決 GrabCut 背景種子點過少時造成的取偽問題。本論文探究影響 GrabCut 效能的主要因素, 更改參數的設置方式, 並增設額外的切割遞迴回合, 提出三種改良的自動切割方法---稱為 K 改進法、 K 遞增法、回合方法。本文選用微軟英國劍橋研究院提供的圖片庫進行切割實驗, 並利用絕對誤差和、傑卡德相似係數及 F 度量等指標評估切割成效, 證實本文方法可以少量背景種子點自動切割, 且切割效果佳, 執行速度也較快速, 因此認定本論文的三種切割方法比 GrabCut 佳。

Image segmentation plays an important role in computer vision.The purpose of image segmentation is to partition an image into meaningful regions with homogeneous properties automatically.Lots of literature reviews show that GrabCut is a better method in recent years.GrabCut is a foreground extraction method which uses graph algorithm to compute its energy function.GrabCut divides an image into foreground and background with high accuracy.It is particularly noteworthy that GrabCut defined the energy function in terms of region and boundary properties which is in accord with human visual perception.However, GrabCut is a semi-automatic segmentation method.It is not convenient for users because they must choose some background seed points before using GrabCut to segment images.In this thesis, we propose three automatic segmentation methods with good accuracy rate by adjusting GrabCut's architectures and parameters.These three methods are called ``K-Improvement'', ``K-Ascending'', and ``Epoch-Method''.The experiments are conducted on Microsoft Research Cambridge's database.We use sum of absolute difference, Jaccard similarity coefficient, and F-measure to evaluate the result.The experiment results show that the proposed methods have higher performance than CrabCut.