Summary: | 碩士 === 國立臺灣科技大學 === 資訊工程系 === 100 === The SIFT have been widely used in the image recognition applications. The feature keypoints extracted by SIFT are invariant to image scaling and rotation. But we need a lot of storage to store the keypoint descriptors with 128 dimensions. A method was proposed that using the Principal Component Analysis (PCA) to reduce the dimension to 36.
Observing that the PCA-SIFT keypoint descriptors is in normal distribution, we propose to simplify the descriptors using binarization and store the binarized descriptors by perfect hash to speed up the query time. However, to improve the accuracy after binarization we identify several weaknesses after applying the perfect hash technique. Through extensive experiments, the results suggest that the robustness of the SIFT descriptors are greatly affected by the angle of objects to be recognized, the lighting directions, and the size of the input images. Therefore, we conjecture that the applications of using SIFT for object recognition in 3D environment are limited and should be designed under scrutiny.
|