基於合作賽局的機器學習式前景分割方法

A Machine Learning Figure-ground Segmentation Method Based on Cooperative Game

蔡念穎(Nian-Ying Tsai)

指導教授 ： 張隆紋

國立清華大學/電機資訊學院/資訊工程學系/碩士(2013年)

摘要

影像分割技術是影像處理領域中的一個重要且具挑戰性的任務，並且在最近幾年被廣泛討論著。影像前景物件分割的主要目標是於一張影像中，將其前景物件從背景中分割出來。但是，要從一張影像中定義出前景物件區域並不是一件簡單的任務。在之前，前景物件分割就已經被一些互動式的分割技術成功解決，然而分割的準確度以及便利性不盡理想。不同於先前的方法，在這篇論文中我們提出了一個機器學習式(使用Support Vector Machine分類器)的影像前景物件分割方法來從背景中提取出影像中的前景物件。
此外，為了提升物件分割的準確率，在學習的步驟我們還運用了一個由Lloyd Shapley所提出的合作賽局理論來評估影像中各種特徵的重要性。在這個賽局中，每一種特徵就代表一個理性的玩家，而特徵的重要性則代表每位玩家對於整個賽局個別的貢獻度。根據我們的實驗結果，相較於其他現有方法，我們所提出的方法在Oxford Flowers 17 和 Caltech-UCSD Birds-200兩個開放式的資料集都有相當不錯的表現。

關鍵字

合作賽局 ； 影像分割 ； 機器學習

並列摘要

Image segmentation is an important and challenging task in image processing, and it is widely discussed in recent years. The main goal of figure-ground image segmentation is to separate foreground objects from their background. But, it is not a simple task to defining the foreground object sections from background in an image. Before, figure-ground segmentation has been addressed successfully by interactive segmentation works. However, it is not an ideal method in accuracy and convenience. Unlike previous methods, in this paper, we present a novel method for figure-ground segmentation with machine learning Mechanism (SVM classifier) to separate the foreground objects from background.
Furthermore, in order to improve the accuracy of figure-ground segmentation, we also use a cooperative game theory which proposed by Lloyd Shapley to estimate the weight of image features in the training step. In this game, each image feature represents a rational player, and the weight of image features represents the contribution of each player. According to our experiment result, our approach obtains very competitive results on Oxford Flowers 17 and Caltech-UCSD Birds-200 data sets in comparison with other state-of-the-art techniques.

並列關鍵字

cooperative game ； image segmentation ； machine learning

參考文獻

[2]Belongie, Serge, Greg Mori, and Jitendra Malik. "Matching with shape contexts." Statistics and Analysis of Shapes. Birkhäuser Boston, 2006, p.81-105.

[3]Tu, Zhuowen, et al. "Image parsing: Unifying segmentation, detection, and recognition." International Journal of Computer Vision. IJCV, 2005, 63.2 : p.113-140.

[4]Jiang, Hao. "Human Pose Estimation Using Consistent Max Covering." Pattern Analysis and Machine Intelligence. IEEE Transactions on, 2011, 33.9: p.1911-1918.

[5]Blake, Andrew, et al. "Interactive image segmentation using an adaptive GMMRF model." Computer Vision-ECCV 2004. Springer Berlin Heidelberg, 2004, p.428-441.

[9]Shapley, Lloyd S. "A value for n-person games." RAND CORP SANTA MONICA CA, 1952, No. RAND-P-295.