Abstract
Content moderation systems may face challenges in consistently identifying images that are part of the same series as previously removed content, as some automated methods may be configured to primarily detect near-duplicates. Systems and methods are described that may address this by comparing a newly submitted image against a complainant's historical set of removed images. An approach can extract and compare both global feature vectors to measure semantic similarity, for example, using a cosine similarity score, and local feature keypoints to determine geometric consistency, for instance, through a geometric verification process. The resulting similarity signals can be assessed against configurable thresholds. When a new image is determined to meet the similarity criteria, the system may identify it as belonging to the same series, which can facilitate automated actions that supplement manual review and may improve the consistency of policy application.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
N/A, "Automated Identification of Imagery Using Global and Local Visual Features", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9601