Benchmarking unsupervised near-duplicate image detection

作者：

Highlights：

• Unsupervised near-duplicate image detection requires high specificity up to 10−6–10−9.

• Empirical comparison of CNN-based descriptors for near-duplicate image detection.

• Validated, principled methodology to estimate sensitivity and estimate false alarms.

• Fine-tuning CNNs for retrieval is beneficial but may suffer in specificity.

• New set of annotations released for near-duplicate detection benchmarking.

摘要

•Unsupervised near-duplicate image detection requires high specificity up to 10−6–10−9.•Empirical comparison of CNN-based descriptors for near-duplicate image detection.•Validated, principled methodology to estimate sensitivity and estimate false alarms.•Fine-tuning CNNs for retrieval is beneficial but may suffer in specificity.•New set of annotations released for near-duplicate detection benchmarking.

论文关键词：Near-duplicate detection,Convolutional neural networks,Instance-level retrieval,Unsupervised detection,Performance analysis,Image forensics

论文评审过程：Received 6 February 2019, Revised 15 April 2019, Accepted 7 May 2019, Available online 8 May 2019, Version of Record 15 June 2019.

论文官网地址：https://doi.org/10.1016/j.eswa.2019.05.002