A brand-neutral logo detection and redaction with transformer detectors

In today’s digital landscape, documents often contain confidential information such as company logos, which, if exposed, may lead to privacy breaches and identity theft. These logos can inadvertently reveal sensitive information, making their detection and redaction essential for the preservation of privacy. Using vision transformer-based techniques, the aim of this research is to develop an automated redaction system to efficiently detect and redact logos. The proposed model used the LogoDet-3K object-based dataset including 3000 classes and 158K logos for training and evaluation. The research also handles the challenge of new, unseen logos by re-annotating the dataset to perform a brand-neutral redaction. In order to check the robustness of the proposed approach, performance was also evaluated on two other standard datasets namely, MS-COCO and PASCAL VOC. The performance of these models is evaluated using the COCO metric, which ensures robust detection of unseen confidential information. The state-of-the-art models have been outperformed by the proposed approach using detection transformer DETR, achieving a mean average precision (mAP) of 0.341 and an average recall (AR) of 0.745 on LOGODet-3K dataset. mAP score of 0.51 and 0.62 has been achieved on MS-COCO and PASCAL-VOC. The proposed pipeline offers a scalable, automated solution for document anonymization in real-world environments.