Research & Publications

An efficient brand-neutral Logo Redaction technique using Detection Transformers

Published paper · Article on Springer | June 2026

• Developed a brand-neutral deep learning pipeline to detect and classify logos in financial documents using DETR, YOLOv5, and DINO on the LogoDet3K dataset.
• Outperformed state-of-the-art baselines on LogoDet3K.
• Performance: mAP = 0.341, AR = 0.745 (robust on unseen data).