Unordered multiple image matching by using descriptor clustering

Abstract

Multiple image matching is an important task in photogrammetry for tie point measurement in block adjustment or generation of point clouds. Numerous algorithms can be found in the literatures. SIFT algorithm, its extracted keypoint descriptor with 128-dimensional vector consisted of the gradient statistics, is developed by Lowe (2004) and becoming a popular method for robust image matching approach by using two images for the decades. When it comes to multiple image matching, the keypoint descriptor is possibly useful. SIFT descriptors of keypoints in all images has been employed to cluster the keypoints to establish the relationship of adjacent images (Chen & Chio, 2013). Therefore, it is possible to use the descriptor clustering for multiple image matching by assuming that keypoint descriptors of the same object point from the different images will be clustered in the descriptor space. In other words, when all keypoint descriptors describe the same object point, the keypoint descriptors will be aggregated in descriptor space within a small region. In this study, adaptive K-means will be used to find all the possible clusters of keypoint descriptors automatically without any initial data. After all clusters are obtained, multiple image matching is also finished. The tests will be performed to prove the proposed idea is able to cluster the descriptors and to perform image matching for keypoints among multiple images successfully.