使用邊緣偵測與基於路徑規劃的方法解決影像分割中需要使用者定義標籤的問題

Using boundary detection and path-based method to solve labeling problem in image segmentation

蘇貫中(Guan-Chung- Su)

指導教授 ： 張隆紋

國立清華大學/電機資訊學院/資訊系統與應用研究所/碩士(2014年)

摘要

在此篇論文當中，我們根據影像中邊緣偵測的方法並結合了賽局理論在影像處理中分群的概念，衍生出一套在影像中尋找數個物件邊緣，並連結起來成為路徑，最後將路徑視為圖片中可能物件之輪廓的方法。並且隨著應用的方法不同，可以將中間過程產生的結果加以調整成不同物件切割方法的輸入標籤，同時我們保持使用我們的方法能夠不用經過使用者的操作來完成，而且還能處理多個的標籤的情形，藉此達到解決目前大多數方法需要使用者定義標籤的問題。
在論文的實驗結果中，則是以Spectral Matting為範例作物件切割的方法，我們利用最終路經包圍的區域，從路徑上的每一線段選擇周遭灰階值最亮與最暗的兩點RGB顏色值，在經過一些計算後能夠得到使用者可能劃分前景與背景標籤的區域，在將這筆結果輸入到spectral matting方法中，得到一個能令人滿意的結果。

關鍵字

影像分割 ； 賽局理論

並列摘要

In this paper, we proposed a method that based on image boundary detection and game theory for multiple labeling in image segmentation. We develop an approach to find the boundaries of multiple objects in the image and then connect the edge segment into a whole path that can be seen as the contour of the object. Furthermore, we can adapt our process into various types of scribbles depending on user’s application. The method is unsupervised and can solve multiple labeling for image segmentation.
In our experiment, we use Spectral Matting as demonstration of our method. We use the area that each path is located and pick the brightest and darkest pixel’s RGB value from the path’s segment patch. Then we can construct the probability map for user’s foreground and background labels. We choose the pixel which bypass the threshold as foreground label; and let the pixels which under the threshold become background label. We use the label result as input scribbles of Spectral Matting to obtain a proper image alpha map.
We compare our result generated from supervised version of Spectral Matting with the unsupervised ones. We can find multiple subtle objects in an image. Nevertheless, the unsupervised Spectral Matting tends to obtain a flat area or incomplete object boundary. Therefore, our method ensures a better result in an unsupervised favor and just cost a little computational time.

參考文獻

[2] J. Shi and J. Malik. “ Normalized cuts and image segmentation. “ IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):888-905, 2000.

[3] Y Weiss. “Segmentation using eigenvectors: a unifying view. “ In Proceedings of the Seventh IEEE International Conference on Computer Vision, volume 2, pages 975-982, Kerkyra, Greece, 20-27 September 1999.

[5] Lim, Joseph J., C. Lawrence Zitnick, and Piotr Dollár. "Sketch tokens: A learned mid-level representation for contour and object detection." Computer Vision and Pattern Recognition (CVPR), 2013 IEEE Conference on. IEEE, 2013.

[6] Fischer, Bernd, and Joachim M. Buhmann. "Path-based clustering for grouping of smooth curves and texture segmentation." Pattern Analysis and Machine Intelligence, IEEE Transactions on 25.4 (2003): 513-518.

[8] D. Comaniciu and P. Meer. “Mean shift: a robust approach toward feature space analysis. “ IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(5):603-619, 2002.