Aggregation-Based Probing for Large-Scale Duplicate Image Detection.

Identifying visually duplicate images is a prerequisite for a broad range of tasks in image retrieval and mining, thus attracts heavy research interests. Many efficient and precise algorithms are proposed. However, compared to the performance duplicate text detection, the recall for duplicate image detection is relatively low, which means that many duplicate images are left undetected. In this paper, we focus on improving recall while preserving high precision. We exploit hash code representation of images and present a probing based algorithm to increase the recall. Different from state-of-the-art probing methods in image search, multiple probing sequences exist in duplicate image detection task. To merge multiple probing sequences, we design an unsupervised score-based aggregation algorithm. The experimental results on a large scale data set show that precision is preserved and the recall is increased. Furthermore, our algorithm on aggregating multiple probing sequences is proved to be stable. © 2013 Springer-Verlag.