Pattern-Expandable Image Copy Detection

Open-world visual recognition aims to empower models to identify objects in real-world settings, particularly when they encounter domains or categories that are not included in the training dataset. This paper proposes a specific open-world visual recognition task, i.e. Pattern-Expandable Image Copy Detection (PE-ICD). In realistic scenarios, the continuous emergence of novel tampering patterns necessitates fast upgrades to the ICD system to prevent confusion in already-trained models. Therefore, our PE-ICD focuses on two aspects, i.e., rehearsal-free upgrade and backward-compatible deployment: (1) The rehearsal-free upgrade utilizes only the new patterns to save time, as re-training on the old patterns can be very time-consuming. (2) The backward-compatible deployment allows for comparing the updated query features against the outdated gallery features, thereby avoiding the need to re-extract features for the extensively large gallery. To lay the foundation for PE-ICD research, we construct the first regulated pattern set, CrossPattern, and propose Pattern Stripping (P-Strip). CrossPattern regulates both base and novel patterns during the initial training and subsequent upgrades. Given a query, our P-Strip separates the tamper patterns by decomposing it into an image feature and multiple pattern features. The advantage of P-Strip is that we can easily introduce new pattern features with minimal impact on the image feature and previously seen pattern features. Experimental results show that P-Strip supports both rehearsal-free upgrading and backward compatibility. Our code is publicly available at https://github.com/WangWenhao0716/PEICD.