then compute those on a pyramid scale space built without smoothing. Moreover, we apply CPU-based parallel processing, SIMD (Single Instruction Multiple Data) in order to speed up the keypiont detection without additional requirement of hardware resources.
Second, we present a floating-point descriptor which is based on intensity rank orders. Intensities have enough distinctiveness, however it could not be used as descriptors directly due to the lack of the robustness for brightness changes. We reinforce the robustness of the intensity by a rank-order normalization. The computational requirement for intensity rank orders is reduced by using the histogram equalization instead of sorting algorithms.
Finally, we propose an intensity-rank-order-based binary descriptor. A previous intensity-rank-order-based binary descriptor shows poor recognition rates and high computational complexity. Fixed threshold for intensity rank orders on a rotation-normalized image patch causes poor recognition rates. The proposed binary descriptor resolves the recognition problem by adaptive binning, and has lower computational complexity than intensity-difference-based binary descriptors due to the histogram equalization and the lookup table.
We evaluated the performance of the proposed local invariant features extensively using the publicly available evaluation toolkits and databases. The experiment results of the proposed methods are compared to those of state-of-the-art methods, and show high performance. Real image sequences, which are captured under various image acquisition conditions, and synthetic image sequences, which are generated by finer-level deformations, are used to evaluate the features. Features are compared in terms of the repeatability for detection, and the recall and precision for matching. In order to analyze the computational complexity of each part of the feature, we compared the timings of the detection, the description, and the direction respectively. Moreover, we applied the proposed method to recognition tests to prove the usefulness of the proposed method.; Local invariant feature may be the most efficient method for finding correspondences of image patches. Feature detects regions called keypoint in an image, then encode information of a local neighbor of the keypoint into a vector called descriptor. Correspondences of detected keypoints in two or several images could be found by comparing similarities of descriptors. Correspondences of image regions are very basic and core information for various computer vision techniques, such as image retrieval, object recognition, motion estimation, 3D reconstruction, and SLAM (Simultaneous Localization and Mapping). As a basic element technology, local invariant feature needs to have high performance in terms of the precision and the computational efficiency. In this dissertation, we present a fast and robust local invariant feature using intensity rank order.
First, we introduce a fast and robust method to detect blob-like structures. For that, we approximate the determinant of the Hessian matrix, which is known as the blob measure, by a small number of pixels