Feature Detector Based Image Copy Detection

Image copy detection is a challenging work in the field of computer vision, and it is of significant practical value. Traditional image copy detection methods can be clas-sified into two categories:global feature based methods and local feature based ones. Global feature based methods, which is efficient to calculate, can only be applied to limited range of transformation, and is hard to deal with the complicated transforma-tion. On the contrary, local feature based approaches achieve promising accuracy, but suffer from extracting the invariant local features and matching these high-dimensional feature descriptors, which bring in a considerable computation cost.Based on these facts, a feature detector based algorithm is proposed to detect copy images in this thesis. We significantly reduce the computational cost by avoiding the time-consuming procedure, local feature description, while maintain the accuracy of local feature based methods. After the maximal stable extremal regions (MSER) is extracted in the algorithm, the simple and cheap geometric parameters and local color histograms is employed to form the initial correspondence of MSER regions. Then based on the combinations of these correspondences, the algorithm accumulates the affine transformation matrices that may exist between the image pairs. A clustering stage extracts potential transformation from the affine transformation matrices, fol-lowed by a verification step.In order to further accelerate the execution of the algorithm, we develop several effective pruning strategies to avoid unnecessary computation by exploiting the proper-ties of the transformation and the actual situation in the topic. Moreover, we adjust the order of varied stages and apply appropriate parallelizing optimization strategies aim-ing at processing large scale of image copy detection task, which lead to a promotion in such circumstance.Finally, experiments are conducted on standard dataset to verify the efficiency and effectiveness of proposed algorithm. At the same time, we discuss about the scalability of our algorithm in terms of the selection of local feature detectors and distance metric through detailed experiments.