Abstract
本研究討論影像比對之搜尋策略，使比對可以不受物體幾何變化之影響。幾何變化會造成非常大之搜尋空間，如果搜尋空間沒有縮減，則在比對前，必須抽取出不變性特徵。不過，傳統的方法僅取出部份的影像資訊，其比對效率比用影像本身做比對來的差。因此本篇論文的主要目的是提出適當的搜尋策略使得影像比對系統可以在縮減之搜尋空間找到可能最佳之比對。 鑑於一般之應用，本篇論文主要關切之議題分為靜態及動態之影像比對，靜態影像比對指兩張影像之比對，動態影像比對表示影像序列中鄰接兩張影像之比對。雖然靜態影像的方法可以應用於影像序列中鄰接兩張影像之比對，但是當考慮動態影像比對的計算量及其特性時，有必要另行開發一個有效率的方法。 靜態影像比對之搜尋策略是基於標定之觀念，對待測影像用相似轉換可縮減待測影像與參考影像之幾何差異。轉換參數是用控制點所估計出來的，這些自動配對的控制點是影像中特定灰階區域的中心，如此，搜尋空間可以大量的縮減。所得之殘餘幾何差異比用正規化的方法小， 因為正規化的方法是用影像動量計算正規化之角度和比例。經由本質不變性特性之分析，僅需準備少數之參考樣板影像即可處理轉換後待測影像之殘餘幾何變化。結果，簡單之相似性量測可以被直接應用，無需比例及旋轉不變性特徵之抽取，更進一步地，控制點集合可以用來做初步之影像比對，如此，如果點樣型比對之結果可令人置信為正確，則某些相似性量測就可以省略，這樣可以節省許多二維之影像相關性計算。兩張影像比對之延伸是影像序列中之比對，用傳統的單一樣板影像比對系統，當物體旋轉或背景有雜亂之物時，會引起相當大之搜尋空間，如此會造成可觀之誤差累積，除此之外，不變性特徵抽取對於動態影像比對是不適宜實行的，因此，雙樣版影像搜尋策略被提出來用在動態影像比對，以尋找連續兩張影像間物體可能最佳之比對。在這個雙樣版影像搜尋策略中， 目前之影像分別與前一張及前第二張影像做比對。其優點是無明顯之計算增加，因為兩個影像比對可以平行處理，而且無需不變性特徵之抽取。用一個不確定性模型，此二個相關性計算之結果可以被綜合分析，結果，搜尋空間大幅減少，物體之最佳可能比對處可以較輕易被找到。兩個綜合分析之方法被提出：貝氏估計法及模糊推斷法。在論文中將會展示，所提出的雙樣版影像搜尋策略優於傳統的單一樣板影像比對系統，並展現壓制比對錯誤累積之能力。This study discusses the search strategies for image matchingunder object'sgeometric transformations, e.g., scaling,rotation, and translation, etc. Itis noted that geometricvariations will cause a tremendous search space. Ifthe searchspace is not reduced, invariant features must be extractedbeforematching. However, conventional methods merelyextract partial imageinformation to achieve invariance.That will result in worse matchingperformance than theresults by using original images for matching. Hence, themainpurpose of this dissertation is to propose matching systemsthat cansearch for the possible best match of objects in a muchreduced search spaceby means of an appropriate searchstrategy. In view of the generalapplications, the mainconcerned issues in this study include static imagematchingand dynamic image matching. Static image matching denotesthematching between two still images. Dynamic image matchingrepresents thematching in an image sequence, in whichconsecutive two images are used formatching. Although themethods for static image matching might also beutilizedfor matching consecutive two images in an image sequence, it isstillnecessary to invent an efficient approach fordynamic matching whenconsidering the computational cost andthe characteristics of dynamic imagematching. The searchstrategy for static image matching is based ontheregistration concept, by which the geometric discrepancybetween the testpattern and the related reference patterncan be reduced by applyingsimilarity transformation tothe test pattern with the transformationparametersestimated from the control points. The automatically pairedcontrolpoints are the centroids of desired regions in patternswith specific graylevel distributions. Thereby, the searchspace can then be much reduced. Theresidual geometricvariations resulted from the proposed method are smallerthanthose resulted from image normalization method whichutilizes imagemoments to evaluate normalization angle andscaling. Through the analysis ofintrinsic invarianceproperties, only a small number of reference templatesarerequired to deal with the residual geometric variations of thetransformedtest pattern. Consequently, a simple similaritymeasure (correlation) can thenbe directly employed withoutany scale and rotation invariant featuresextraction.Furthermore, the control point sets can be utilized as theinitialmathing of two image patterns. As a result, somesimilarity measure can thenbe omitted if the point patternmatching reports a correct and convincingmatch of the twopatterns. That will save much time for two dimensionalimagecorrelation. The extension of two images' matching is thematching in an imagesequence. Conventional single-templatecorrelator tends to bring about aconsiderable amount ofwalk-off errors due to the clutter effect or that theobject inthe search scene varies greatly in orientation. That willcause aconsiderable amount of search space as stated in staticimage matching. Inaddition, the invariant features extractionis not feasible in the dynamicimage matching. Hence, a dual-template search strategy for dynamic imagematching isproposed to seek the possible best match of objects betweentwoconsecutive images. In this dual-template searchstrategy, the AOI of thecurrent frame is correlated with the``precedent first'' template and the``precedent second''one for seeking the possible best match of objects. Ithas theadvantage that no significant increase of computational burdenof thesystem is needed since these two correlations can beimplemented in a parallelmanner and no complicated invariantfeatures extraction is required. By meansof an uncertaintymodel of correlation peaks, the results of thedualcorrelations are analyzed and combined within theintersection of two circulargates determined by the model.Therefore, the search space is diminished andthe possible bestmatch of objects in consecutive frames can then beeasilysearched. Two techniques are proposed for thiscorrelation data fusion andoptimization: by probabilisticBayesian estimation or by fuzzy inference. Itwill be shown thatthe proposed dual-template search strategy outperformstheconventional single-template correlators implemented ineither matched filteror circular harmonic filter. In addition,the proposed method exhibits thecapability of walk-offsuppression.