物件辨識的方式有很多，其中常被人使用的方式為偵測物件特徵點然後對特徵點進行比對。然而影像常常會有旋轉、縮放或是平移的變動，因此在偵測特徵點的同時會受到這些變動的影響。本研究的目的在於找尋物件或是影像的特徵點及其特徵向量並且具有強健性，在影像經過變動之後仍然具有不變性。 本研究利用DWT找出特徵點，接著使用log-polar轉換使特徵具有角度的不變性，利用亮度的差值決定特徵向量的內容以抵抗亮度的改變，最後利用幾何學相似三角形的原理提升比對的正確率。 在實驗中與CCH[1]做比較，確實提昇了Scaling不變性的效果，另外對於亮度變化以及模糊化也有不錯的表現，此外其他的實驗和CCH[1]有著類似的結果。在時間上，跟CCH相差甚少，也就是說相較於SITF[3]快了近兩倍之多。 Feature points’ matching is a popular method in dealing with object recognition problems. However, variations of images, such as shift, rotation, and scaling, influence the matching correctness. Therefore, a feature point matching system with distinctive and invariant feature point detector as well as robust description mechanism becomes the main challenge of this issue. We use discrete wavelet transform (DWT) and accumulated map to detect feature points which are local maximum points on the accumulated map. DWT calculation is very efficient comparing to that of Harris corner detection or Difference of Gaussian (DoG) proposed by Lowe. Besides, feature points detected by DWT are located more evenly on texture area unlike those detected by Harris’ are clustered on corners. To be scale invariant, the dominate scale (DS) is determined for each feature point. According to the DS of a feature point, an appropriate size of region centered at this feature point is transformed to log-polar coordinate system to improve the rotation and scale invariance. A descriptor of dimension 32 is made of the contrast information to enhance the illumination robustness. Finally, in matching stage, a geometry relation is adopted to improve the matching accuracy. Comparing to existing methods, the proposed algorithm has better performance especially in scale invariance and robustness to blurring effect.
