運用超像素與圖形理論的改良式影像分割技術及其在顯著影像偵測上的應用

影像分割是在電腦視覺和影像處理領域中的一個基本問題。雖然本主題已經被研究了許多年，它仍然是一項具有挑戰性的任務。近年來關於超像素(superpixel)的研究有很大的進展，並且這種新技術可以使傳統的影像分割演算法具有更高的效率和更好的性能。另一方面，顯著影像偵測的研究是影像處理的一個新的課題，其表現通常與使用的影像分割技術有深度相關。在這篇論文中，我們提出了兩種演算法分別針對影像分割及顯著影像偵測。對於影像分割，我們基於採用超像素的圖形表示，提出一種有效的影像分割演算法。該演算法採用了包含SLIC 超像素，5-D譜聚類(spectral clustering)和邊界專注(boundary-focused)的區域合併等技術。利用SLIC超像素，原始影像分割問題被轉化為超像素分類問題。這使得該演算法比基於像素的分割演算法能有更佳的執行效率。利用5-D譜聚類和邊界集中的區域合併法，位置資訊可用於集群且區域合併的門檻值可根據影像作自適應調整。這些技術使分割結果更符合人類感知。依據在柏克萊分割數據庫的結果顯示，我們提出的方法優於目前已知最先進的方法。對於顯著影像偵測，我們提出了一種非常有效的顯著影像偵測演算法。我們的演算法主要是基於以下兩個步驟。首先，離散餘弦變換（discrete cosine transform）用於產生塊單位(block-wise)的顯著影像圖。然後，基於超像素的影像分割演算法被應用以得到像素單位(pixel-wise)的顯著影像圖。由於離散餘弦變換係數可以反映在頻域的每個塊的顏色特徵且超像素可以很好地保留物體的邊界，這兩種技術可以大幅改善顯著影像偵測演算法的性能。根據在具有1000張影像的影像數據庫上進行的實驗結果，我們所提出的方法可以非常精確地提取顯著區域，並優於所有現有的顯著影像偵測方法。

關鍵字

影像分割；顯著影像偵測；超像素；電腦視覺

並列摘要

Image segmentation is a fundamental problem in computer vision and image processing. Though this topic has been researched for many years, it is still a challenging task. Recently, the researches of superpixels have great improvement. This new technique makes the traditional segmentation algorithms more efficient and has better performances. On the other hand, the saliency detection is another new topic of image processing and its performance usually closely related to the segmentation techniques we used. In this thesis, we propose two algorithms for image segmentation and saliency detection, respectively. For image segmentation, an effective graph-based image segmentation algorithm using the superpixel-based graph representation is introduced. The techniques of SLIC superpixels, 5-D spectral clustering, and boundary-focused region merging are adopted in the proposed algorithm. With SLIC superpixels, the original image segmentation problem is transformed into the superpixel labeling problem. It makes the proposed algorithm more efficient than pixel-based segmentation algorithms. With the proposed methods of 5-D spectral clustering and boundary-focused region merging, the position information is considered for clustering and the threshold for region merging can be adaptive. These techniques make the segmentation result more consistent with human perception. The simulations on the Berkeley segmentation database show that our proposed method outperforms state-of-the-art methods. For saliency detection, a very effective saliency detection algorithm is proposed. Our algorithm is mainly based on two new techniques. First, the discrete cosine transform (DCT) is used for constructing the block-wise saliency map. Then, the superpixel-based segmentation is applied. Since DCT coefficients can reflect the color features of each block in the frequency domain and superpixels can well preserve object boundaries, with these two techniques, the performance of saliency detection can be significantly improved. The simulations performed on a database of 1000 images with human-marked ground truths show that our proposed method can extract the salient region very accurately and outperforms all of the existing saliency detection methods.

並列關鍵字

Image segmentation ； Saliency detection ； Superpixels ； Computer vision

參考文獻

A. Image Segmentation

[1] X. Ren and J. Malik, “Learning a classification model for segmentation,” In ICCV, pp. 10-17, 2003.

[2] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and S. Süsstrunk, “SLIC superpixels compared to state-of-the-art superpixel methods,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 34, no. 11, pp. 2274 - 2282, May 2012.

[3] D. Martin, C. Fowlkes, D. Tal, and J. Malik, “A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics,” In ICCV, pp.416-423, 2001.

[4] T. Kim and K. Lee, “Learning full pairwise affinities for spectral segmentation,” In CVPR, pp. 2101-2108, 2010.

國際替代計量

運用超像素與圖形理論的改良式影像分割技術及其在顯著影像偵測上的應用

全文下載

主題瀏覽