針對巨量影像檢索使用本文資訊之區域描述子配對法

Abstract

基於MPEG壓縮描述子用於視覺化檢索(CDVS)之架構，本論文提出一個利用本文資訊，濾掉特徵點配對中錯誤配對的方法。在區域描述子量化以及配對缺乏幾何資訊下，特徵點配對很容易產生非常多的錯誤，影響幾何驗證的判斷。因此，本論文嘗試結合了半區域演算法和虛弱幾何一致性兩方法，有效的將錯誤的配對濾除，進而減少幾何驗證的誤判率。為了減輕本文確認自身複雜度，本論文提出了以本文資訊為基礎之適配度檢定(WGC Goodness-of-fit Test)，有效將近一半的雜訊影像省略幾何驗證的步驟。最後，在觀察到特徵點配對才是檢索中最複雜之下，本論文亦提出兩種以雜湊為基礎的特徵點配對方法，有效地降低整體檢索架構時間，並維持檢索精確度。實驗結果發現在mAP上相對於CDVS有1.0-2.1%的進步，而就檢索時間的分析上，本文資訊適配度檢定可以節省約4-14%的時間，而雜湊特徵點配對可以節省將近40-50%的時間。

A key problem in MPEG-7 Compact Descriptors for Visual Search (CDVS) framework is the ambiguity of feature matching. To alleviate it, a new contextual verification scheme is introduced in CDVS by combining semi-local constraints and weak geometric consistency check. To mitigate time complexity incurred by the contextual verification, we propose a goodness-of-fit test based on the features’ orientation, which is motivated by the CDVS goodness-of-fit test. Moreover, we propose two hash-based feature matching schemes to speed up the feature matching process, which is found to be the most time-consuming process in the current CDVS framework. Experimental results show that the contextual verification offers 1.0-2.1% mAP improvements over CDVS. For time reduction experiments, 4-14% and 40-50% time savings are achieved by our goodness-of-fit test and hash-based feature matching scheme.