Abstract
Image copy detection (ICD) aims to determine whether a query image is an edited copy of any image from a reference set. Currently, public benchmarks for ICD overlook a critical real-world challenge: distraction from hard negative queries. These are images that are inherently similar to reference images but are not edited copies, leading to high false-positive rates. This paper introduces the NDEC (Negative Distractor for Edited Copy) benchmark, featuring over 124,000 hard negative pairs. Furthermore, we identify a fundamental conflict between symmetric metric learning and the asymmetric nature of image editing (e.g., cropping). We propose Asymmetrical-Similarity Learning (ASL) to address this, allowing for directional similarity. Experimental results demonstrate that ASL significantly outperforms state-of-the-art methods on the NDEC benchmark.
Introduction
Practical Image Copy Detection (ICD) is essential for copyright protection and content management. However, existing systems struggle with "hard negatives"—images that look nearly identical but are distinct captures. Moreover, standard retrieval models rely on symmetric distance metrics, which fail to account for the unidirectional nature of many edits. For instance, while a cropped image is a copy of the original, the original is not a copy of the crop. This paper addresses these gaps through a new benchmark and a novel learning framework.
The NDEC Benchmark
To better reflect real-world conditions, we developed the Negative Distractor for Edited Copy (NDEC) dataset. Building upon the DISC21 dataset, we manually annotated and added 100,000 hard negative pairs to the training set and 24,252 to the test set. These distractors are designed to confuse models that rely solely on visual similarity without understanding the underlying relationship between images.
Asymmetrical-Similarity Learning (ASL)
To resolve the symmetry conflict in metric learning, we propose Asymmetrical-Similarity Learning (ASL). The core intuition is that an original reference image typically contains more information than its edited copy. We utilize the feature norm (the magnitude of the feature vector) as a proxy for this information density.
The ASL Mechanism
We define a similarity ratio $R(x_i \to x_j) = |x_i| / |x_j|$. During training, the model is optimized to ensure that for a reference-to-copy pair, this ratio is greater than 1. By integrating this norm-aware loss with standard distance-based losses (such as CosFace), the model learns to distinguish the "source" from the "copy" based on vector magnitude, effectively breaking the symmetry of traditional Euclidean or Cosine distances.
Experimental Results
Our evaluation on the NDEC benchmark reveals that existing state-of-the-art (SOTA) models perform poorly when faced with hard negatives. For example, the performance of the D2LV model dropped from ~90% $\mu AP$ on DISC21 to ~59% on NDEC.
In contrast, ASL achieved a superior 64.09% $\mu AP$ on NDEC. Most notably, ASL provided a massive boost in precision, jumping from 45.30% to 73.84% over a baseline model. This confirms that addressing the symmetric-asymmetric conflict is vital for reducing false positives in practical ICD applications.
Conclusion
This work highlights the limitations of current ICD benchmarks and the inherent flaws in applying symmetric metric learning to asymmetric editing processes. By introducing the NDEC benchmark and the ASL method, we provide a more robust framework for detecting image copies in the presence of challenging distractors. Future work will focus on refining the relationship between feature norms and stylistic edits beyond cropping.
Summary updated