運用優角、陰影補償、及顯著區域偵測的改良式影像分割技術

Improved Image Segmentation Techniques Based on Reflex Angles, Shadow Compensation, and Salient Region Detection

蔡佳豪(Chia-Hao Tsai)

指導教授 ： 丁建均

國立臺灣大學/電機資訊學院/電信工程學研究所/碩士(2011年)

https://doi.org/10.6342/NTU.2011.01360

摘要

影像切割在影像處理為一大挑戰工作，在影像分析和圖像識別的前處理極為重要，並決定了分析之最後結果的品質。影像切割是用來把一張影像切成好多不同且非重疊的同質區域，在此同質區域基於不同的條件像是灰階值、顏色或紋理，而組成。因此，我們希望去找出一些新的切割演算法去滿足不同的需求。
在這篇論文中，首先我們簡單介紹快速掃描演算法(fast scanning algorithm)，它的每個像素只需要被處理一次。基於快速掃描演算法，我們改變它的距離區間，並加入了邊緣資訊，然後使原本不能分開的兩個區域，能夠分成兩個不同的區塊。模擬結果顯示我們的演算法使切割結果變的更好。
從人類的觀點，現存的切割演算法之切割結果不是足夠好，或者可以說，它與實物匹配(matching physical objects)能力不好。在此我們結合優角(大於180度的角)運算以及使用簡單方法去找代表點。實驗證明我們的方法改善了現存方法的實物匹配能力。此外，我們演算法的處理時間比使用形態學運算更快。
由於物體表面的明亮度改變，陰影和亮光對於電腦視覺研究者來說，都是極為困難的挑戰。對於很多的影像分析和應用，像是物件抓取和描述來說，陰影干擾是基本要處理的工作。此外，對於影像測量上物體、陰影和鏡面反射的幾何性質變異極大，因此，單一材料反射率的切割也是相當具有挑戰性的問題。
換句話說，對於影像分析來說，陰影切割是一重要步驟。基於快速掃描演算法，我們提出一正規化方法去解決陰影和亮光所產生的問題。模擬結果顯示我們的演算法可以成功幫助我們切割含有陰影的影像。此外，我們演算法對於含有陰影的影像之物體與背景之切割完整性，比其它存在演算法好。
顯著區域偵測被使用在很多方面，像是物件識別、自適性壓縮技術和物件切割。在此論文中，我們介紹一顯著區域偵測的方法，並用山脊分佈分析(Ridge-based Analysis of Distributions (RAD)) 演算法以及快速掃描演算法當作預切割的方法。山脊分佈分析演算法結合物理、特徵方法，以及單一材料反射率的分佈之觀察。
最後，我們利用正確解答的圖形資料庫，並使用不同預切割的方法包括山脊分佈分析演算法、快速掃描演算法，以及平均位移演算法(mean shift algorithm)，去計算精確率(precision)和召回率(recall)。實驗結果證明在精確率接近的情況下，不論使用固定或是適應性參數，使用快速掃描演算法的召回率皆優於山脊分佈分析演算法與平均位移演算法。

關鍵字

影像切割 ； 影像處理 ； 影像分析 ； 圖像識別 ； 邊緣資訊 ； 大於一百八十度的角 ； 形態學 ； 陰影切割 ； 顯著區域偵測 ； 物件識別 ； 自適性壓縮技術 ； 物件切割

並列摘要

Image segmentation is a big challenging task in image processing. Moreover, it isvery important pre-processing in image analysis and pattern recognition, and determinesthe quality of the final result of analysis. A process of partitioning an image into different non-overlapping homogeneous regions is called image segmentation, where
the homogeneous regions may be composed based on different criteria such as gray-level, color or texture. Therefore, we hope to find and propose some new segmentation algorithms for satisfying various requirements.
In the thesis, first of all, we briefly introduce the fast scanning algorithm that eachpixel is processed only once. Based on the fast scanning algorithm, we change its
distance interval and add in the edge information which can make the two original inseparable regions divide into two different regions. The simulation results show that our algorithm makes the segmentation result become be better.
From human’s point of view, the segmentation result of the existing segmentation algorithm is not good enough, or to be precise, its ability for matching physical objects
is not good. Here we combine the reflex angle operation with the method of finding representative points which a simple method. The experimental results show that the
improved method improves the existing segmentation algorithm’s the ability for matching physical objects. Moreover, the processing time of our algorithm is faster than using the morphology operation.
Due to variance in the brightness on the surfaces of the objects under consideration, shadows and highlights present a challenge to the computer vision researchers. Shadows
interfere with fundamental tasks in many image analysis and interpretation applications such as object extraction and description. Furthermore, owing to the great variation in
image measurements caused by the geometry of the object, shadows, and specularities, the segmentation of a single material reflectance is a quite challenging problem.
In a word, shadow segmentation is an important step in image analysis. We propose normalized method to solve the shadow and highlight problem based on the fast scanning algorithm. The simulation results show that our algorithm can successfully help us to segment the shadowed images. Moreover, the completeness of the segmentation results of the object and the background of the shadowed images are
better for our algorithm than the other existing algorithms.
Salient region detection is used in many applications, such as object recognition, adaptive compression, and object segmentation. We introduce a method for salient
region detection in the thesis, and use the Ridge-based Analysis of Distributions (RAD) and the fast scanning algorithm for the pre-segmentation algorithm. The RAD algorithm combines the strengths of physics with feature-based methods, and is based on the
observation that the distribution of single material reflectance.
Finally, we compare the salient object segmentation result of using the RAD and the fast scanning algorithm with the mean shift segmentation algorithm and calculate
the precision and recall by means of the ground truth image database. The experimental results show that no matter using fixed or adaptive parameters, the recall values of using the fast scanning algorithm are better than using the RAD or mean shift segmentation algorithm under the close precision value condition.

並列關鍵字

image segmentation ； image processing ； image analysis ； pattern recognition ； edge information ； reflex angle ； morphology ； shadow segmentation ； salient region detection ； object recognition ； adaptive compression ； object segmentation

參考文獻

A. Thresholding

[1] R. C. Gonzalez, and R. E. Woods, "Chapter 10: Image Segmentation," Digital Image Processing 3rd Ed., pp.738-763, Prentice-Hall, 2008.

[2] S. Arora, J. Acharya, A. Verma, and Prasanta K. Panigrahi, "Multilevel thresholding for image segmentation through a fast statistical recursive algorithm," Pattern Recognition Letters, vol. 29, Issue 2, pp. 119-125, Jan. 2008.

[3] F. Yan, H. Zhang, and C. R. Kubeb, "A multistage adaptive thresholding method," Pattern Recognition Letters, vol. 26, Issue 8, pp. 1183-1191, June 2005.

[4] F. Kurugollu, B. Sankur, and A.E. Harmanci, "Color Image Segmentation Using Histogram Multithresholding and Fusion," Image and Vision Computing, vol. 19, issue 13, pp. 915-928, Nov. 2001.