變形圖樣判別及影像定位之研究

A Study on Deformed Pattern Identification and Image Registration

劉宜生(Yi-Sheng Liu)

指導教授 ： 劉如生 ； 陳淑媛

元智大學/資訊學院/資訊工程學系/博士(2008年)

https://doi.org/10.6838/YZU.2008.00082

摘要

圖樣判別(pattern identification)及影像定位(image registration)是影像處理應用領域中的兩項基礎工作，兩者的目的都是經由偵測(detection)及比對(matching)以找出兩張或多張影像之間的異同。拍攝的影像可能因為使用不同的感測器、或者拍攝時間、狀況、視角、物體等條件的不同，造成影像之間不同程度的差異與變形，執行上述兩項工作所使用的方法，也必須針對不同的應用目的及影像特性做適當的調整。系統對於變形影像的處理能力，正可以考驗系統的強韌度。本研究的目的是為交通標誌及二維電泳圖分析分別設計一個強韌的圖樣判別及影像定位方法。
在本研究的第一部分，將針對比例變化(scaled)及歪曲變形(skewed)的交通標誌提出一個快速而強韌的圖樣判別方法。在標誌偵測階段，首先將輸入的彩色影像進行HSV顏色轉換及取樣。接著使用邊緣追蹤(border tracing)找出影像中與交通標誌外框顏色相同的ROI(Region of Interest)區塊。根據ROI自動調整圓形及三角形樣版的大小，ROI外緣的點也根據ROI的長、寬比例自動調整，再與產生的樣版進行比對。本方法對於有複雜背景的交通標誌亦可便利地進行偵測及驗證。同時，因為ROI形狀的調整與驗證僅須處理ROI邊緣的像素，所以所提的偵測方法，其執行的速度很快。在標誌辨識階段，首先將偵測到的交通標誌縮放至固定大小(96x96像素)，接著根據分佈關聯(layout context)進行直方圖比對(histogram matching)運算，以度量景象中標誌與資料庫中標誌兩者間的相似度，進而完成交通標誌辨識。因為基於分佈關聯直方圖比對具有執行速度快及變形容忍度高的特點，因此，所提的方法有很高的辨識率及執行效率。
在本研究的第二部份，針對二維電泳圖影像分析提出一個強韌的影像定位方法。在斑點偵測階段，所提的方法是先在電泳圖的灰階方向取橫切面，以取得斑點的切片(slice)，將其建成切片樹(slice tree)，再根據切片樹進行斑點偵測。每個斑點在其灰階範圍可以取得一系列的切片，每個切片有各自的灰階值、大小、形狀、中心位置及邊界等相關資訊。如果將這些切片中心點投影到共平面，屬於同一個斑點的切片中心點會落在一個相鄰的範圍，這些中心點的落點分佈與斑點的形狀有關。因為切片包含了斑點的相關資訊，而切片樹又包含所有切片之間的關聯資訊，依據切片樹進行斑點偵測可有效提升偵測的精確度，同時解決現有商業斑點偵測軟體採用分水嶺法(watershed)所引發的過度切割問題。
在斑點比對階段，則以分佈關聯(layout context)及鬆弛運算(relaxation)為基礎，使用疊代迴圈來處理電泳圖的非線性變形，整個程序又可分為估算(estimation)、修正(refinement)、及轉換(transformation)等三個階段。在估算階段，利用每個斑點的分佈關聯來計算每兩個斑點間的配對成本(matching cost)，再根據配對成本初始化鬆弛運算所使用的配對機率矩陣(matching probability matrix)。在修正階段，使用鬆弛運算反覆地修正配對機率矩陣，使斑點的配對關係達到整體一致性。在轉換階段，從配對的斑點中慎選出適當的控制點，以建構兩張電泳圖之間的細版曲線(thin-plate spline, TPS)轉換參數。將其中一張電泳圖利用TPS進行幾何轉換，使兩張電泳圖配對的斑點可以互相更靠近。藉著重複執行「估算-修正-轉換」迴圈，讓電泳圖的斑點配對逐漸達到穩定狀態，同時因為所提方法分別在修正與轉換階段，納入慎選配對機率計算之標記點及轉換參數計算之控制點，故所提方法可得最佳配對結果。
從實驗數據分析，在交通標誌的研究部分，我們的偵測率及辨識率分別可以達到94.2%及91.7%，平均偵測一張640x480影像所花的時間約4-50 ms，而辨識一個交通標誌的時間約需10 ms，證實所提的方法快速而有效。在二維電泳圖的研究部分，實驗結果顯示斑點可以被正確地偵測出來，而所提的電泳圖定位方法也能精確配對，同時不論是在斑點偵測或比對，就全自動化處理模式而言，所提方法都較商業軟體ImageMaster 2D有較高的準確度。

關鍵字

圖樣識別 ； 圖樣判別 ； 影像定位 ； 交通標誌偵測 ； 交通標誌辨識 ； 二維電泳圖 ； 斑點偵測 ； 斑點比對 ； 分佈關聯 ； 直方圖比對 ； 鬆弛運算 ； 細版曲線轉換

並列摘要

Pattern identification and image registration are essential tasks in the applications of using image processing. Both of their purposes are to find a match or mismatch between two or more images through the detection and matching stages. Since the images may be taken from different sensors, at different time or under different conditions, from different viewpoints, or for different objects, the methodologies used in these tasks should be heavily adapted to the objectives of applications and the characteristics of images. Moreover, the deformation problems will be encountered in the design of these tasks and invoke the difficulties to achieve robustness. Design of robust pattern identification for road signs and image registration for two-dimensional electrophoresis (2-DE) gel analysis is the goal of this study.
In the first part, a fast and robust pattern identification method for scaled and skewed road signs is proposed. In the detection stage, the input color image is first quantized in HSV color model. Border tracing those regions with the same colors as road signs is adopted to find the regions of interest (ROI). ROIs are then automatically adjusted to fit road sign shape models to facilitate detection verification even for scaled and skewed road signs in complicated scenes. Since the ROI adjustment and verification are both performed only on border pixels, the proposed road sign detector is fast. In the recognition stage, the detected road sign is normalized first. Histogram matching based on layout context is then used to measure the similarity between the scene and model road signs to accomplish recognition. Since histogram matching is fast and has high tolerance to distortion and deformation, our method has high recognition accuracy and fast execution speed.
In the second part, a robust image registration for 2-DE gel analysis is proposed. In the spot detection stage, the proposed method takes slices of a gel image in the gray level direction and builds them into a slice tree, which in turn is used to perform spot detection. More specifically, a series of slices of spots can be obtained in the intensity direction. Each slice of a spot has its own features such as size, shape, central point and boundary smoothness. If the central points of slices are projected onto the co-plane, the projected points belonging to the same spot will fall in a neighborhood. The distribution of these projected points vary according to the shape and appearance of the spots in a gel image. Since the information of slices, namely slice context, can be embedded into a slice tree based on which the proposed spot detection can resolve the over-segmentation problem. Over-segmentation is a well-known drawback of Watershed adopted by most commercial spot detection software.
In the spot matching stage, an iterative approach based on layout context and relaxation labeling is proposed to cope with non-linear deformation of gel images. The proposed matching uses an estimation-refinement-transformation strategy. In the estimation phase, layout context for each spot is used to calculate matching cost between each pair of spots. The matching cost is then used to initialize a matching probability matrix for consequent relaxation labeling. In the refinement phase, relaxation labeling is used to iteratively update the matching probability matrix to achieve global consistency. In the transformation phase, control points are subtly selected to calculate the thin-plate spline (TPS) parameters between the gel images. The TPS parameters are then used to transform the spots in a gel image closer to the corresponding spots in the other. The estimation-refinement-transformation loop is executed iteratively until the matching correspondence between spots reach a stable state. Since subtle selection of landmark points for matching probability computation in the refinement phase and control points for transformation parameter determination in the transformation phase is involved in the proposed method, optimal matching results are obtained.
Experimental results show that the detection rate and recognition accuracy of the proposed road sign identification can reach 94.2% and 91.7%, respectively. On an average, it takes only 4-50 and 10 ms for detection and recognition, respectively. Thus, our method is effective, yet efficient. On the other hand, experimental results show that spots can be accurately detected and the spot matching pairs can be accurately identified. Moreover, for the fully automatic mode, the proposed method has higher accuracy than the commercial software ImageMaster 2D both in spot detection and spot matching.

並列關鍵字

Pattern recognition ； pattern identification ； image registration ； road sign detection ； road sign recognition ； 2-DE gel electrophoresis ； spot detection ； spot matching ； layout context ； histogram matching ； relaxation labeling ； thin-plate spline

參考文獻

[1] P. Douville, “Real-time classification of traffic signs,” Real-Time Imaging, vol. 6, pp. 185–193, June 2000.

[2] A. de la Escalera, L. E. Moreno, M. A. Salichs, and J. M. Armingol, “Road traffic sign detection and classification,” IEEE Trans. on Industrial Electronics, vol. 44, pp. 848–859, Dec. 1997.

[3] S. Hsu and C. L. Huang, “Road sign detection and recognition using matching pursuit method,” Image and Vision Computing, vol. 19, pp. 119–129, Feb. 2001.

[4] Y. B. Lauziere, D. Gingras, and F. P. Ferrie, “A model-based road sign identification system,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition, pp. 1163–1170, Dec. 2001.

[5] P. Paclik, J. Novovicova, P. Pudil, and P. Somol, “Road sign classification using Laplace kernel classifier,” Pattern Recognition Letter, vol. 21, pp. 1165–1173, Dec. 2000.