基於強健 GrabCut 與醒目區域偵測的自動化前景選取

Abstract

在基於內容的圖像檢索中,影像切割是十分重要的前處理步驟。針對基於內容的圖像檢索應用, GrabCut 具有良好的切割成效,它是基於能量函數的切割方法,先以像素點間以及像素點與模型間的色彩相似程度定義能量函數,再結合圖論演算法進行切割的方法。然而, GrabCut 屬於半自動切割法,使用者必須標記初始背景像素點,並且不容許標記錯誤。為了達到自動化切割,本研究藉由偵測圖片中的醒目區域取代使用者標記,並提出強健 GrabCut 切割法,使容許部分錯誤標記,此自動切割方法稱為醒目區域標記法。為了降低醒目區域標記法的棄真和取偽錯誤,我們提出邊框背景標記法和二階段簡化背景標記法。本論文使用微軟英國劍橋研究院所提供的圖片庫以及柏克萊切割數據集的圖片進行實驗,並利用 F 度量評估切割成效。實驗結果顯示藉由邊框背景標記法以及二階段簡化背景標記法確實能夠提升自動切割的準確性。

Image segmentation is an important preprocessing step in the content-based image retrieval. For image retrieval applications that are content based, GrabCut performs well in segmentation. GrabCut is an energy-based segmentation, it defines energy functions with similarity of color between pixels and between pixels and models first. It then combines graph algorithms to proceed with the segmentation process. However, GrabCut is a semi-automatic segmentation method, therefore the user must label some background pixels before using it to segment images, incorrect pixels labeling are also not allowed. In order to achieve automatic segmentation, this research uses detection of salient areas in the images to replace user labels. We then propose a Robust GrabCut segmentation method to allow incorrect pixel labelings. This automatic segmentation method is called “Salient Region Labeling”. In order to reduce Salient Region Labeling’s error in false negatives and false positives, we propose two labeling methods. These two methods are called “Background Frame Labeling,and “Two-stage Simplify Background Labeling”. The image database from Microsoft Research Cambridge and BSD500 were used in this experiment, and then the F-measure was used to evaluate the results. The experiment results show that Background Frame Labeling and Two-stage Simplify Background Labeling methods can effectively enhance the accuracy rate of automatic segmentation.