應用改良式直覺角點偵測與二階段角點匹配於物件辨識

Object Recognition Based on Modified Intuitive Corner Detection and Two-Stage Corner Matching

陳金聖(Chin-Sheng Chen) ； 蔡銘富(Ming-Fu Tsai) ； 邱駿展(Chun-Chan Chiu)

《臺北科技大學學報》 42卷2期 (2009/12) Pp. 19-35

https://doi.org/10.29768/JNTUT.200912.0002

摘要

本研究提出一個快速物件辨識的系統，本文辨識系統之架構可分為離線訓練階段與線上操作階段，離線訓練階段進行目標物的角點提取、多解析度樣本取樣，與其特徵描述向量計算，最後建立階層式的物件資料結構；線上操作階段進行輸入影像的角點提取與其特徵描述向量計算、角點匹配、物件識別。因本辨識系統上同時考量穩定性與即時性，所以角點擷取部份是採用改良式直覺角點偵測方式以達到快速提取的效果，特徵描述向量使用SIFT的特徵描述，並透過PCA的方式來降低SIFT描述空間的維度。在進行物件辨識時，運用二階段角點匹配，縮減角點匹配空間來加快比對速度，最後搭配RANSAC去除匹配錯誤的點，達到正確的物件辨識。實驗結果顯示：(1)本文的改良式直覺角點偵測，其重現性優於傳統的直覺角點偵測，但Harris角點偵測之重現性較佳；而角點提取時間略遜於傳統的直覺角點偵測，但遠優於Harris角點偵測。(2)採用二階段角點匹配，在匹配時間上明顯少於全域角點匹配，且辨識效果依舊強健。(3)在物件辨識上，因建立多解析度樣本資料庫，雖使樣本資料增加11倍，但對於物體大小的改變，有較佳的辨識效果。

關鍵字

多解析度樣本 ； 改良式直覺角點偵測 ； 二階段角點匹配

並列摘要

This paper proposes an object recognition algorithm based on modified intuitive corner detection and two-stage corner matching. The object recognition algorithm consists of two phases: the off-line training phase and the on-line operating phase. The critical purpose is to construct template database in the training phase. Firstly, the corners are extracted from the template image by the modified intuitive corner detection. The multi-resolution patches are then applied to create the full scale corners' features. Each corner has its own descriptor based on SIFT and PCA. With this information, the algorithm creates the hierarchical structures of multi-resolution patches to improve the speed of corner matching. In the operating phase, the test images are processed in the same manner mentioned above with single resolution patches, and then the corner will be matched with the multi-resolution patches in the training phase’s database. The two-stage corner matching, coarse and fine matching based on hierarchical structures of corner descriptions appears to reduce the range of patch’s candidates, is then adopted to improve the matching performance. Finally, the RANSAC criterion is applied to reject the remaining outlier. Experimental results show that: (1) Our proposed corner detection algorithm has better repeatability than traditional intuitive corner detection in almost cases. And the computation time for our proposed algorithm is less than Harris corner detection. (2) In the corner matching, the two-stage corner matching presents the similar matching number of corners and less computation time comparing with full searching algorithm. (3) The object recognition based on the multi-resolution patches will be improved when the object undergoes different scale.

並列關鍵字

Multi-resolution patches ； Modified intuitive corner detection ； Two-stage corner matching