Canopy approach of image clustering based on camera fingerprints

The paper presents a canopy based image clustering algorithm using normalized cross-correlation among the camera fingerprints as a decision criterion. The proposed framework uses two levels of the threshold at two different stages to cluster images based on camera fingerprints. Initially, fingerprints are sorted in descending order of their goodness, and then raw clusters are constructed using a relaxed threshold followed by fine clustering with a hard threshold. The raw and fine clustering process results in non-overlapping clusters, which avoids assigning a fingerprint to multiple raw or fine clusters. The fine clusters are further processed in the attraction phase to improve the cluster's quality at the cost of some computation. The CIC algorithm results in high-quality clusters with a reduced computational cost. The results show that the computational complexity per fingerprint, with respect to the reference complexity of n(n - 1)/2, decreases as the size of the dataset increases. The proposed algorithm also does not suffer from the problem when the number of cameras is larger than the average number of images taken with a camera, i.e., NC >> SC. Hence, the algorithm is suitable for large scale clustering and solving different scenarios of NC >> SC.