Deep Seal Matching: Unleashing the Power of ViT-Triplet and Text Background Removal

In Chinese society, seals are widely used to ensure the authority and validity of critical situations, holding significant importance. Therefore, seal authentication is crucial. Current authentication method relies on manual comparison, which is subjective and requires additional time and labor costs to train authentication personnel. To address these challenges, we propose a novel and highly effective seal matching algorithm. Our algorithm integrates the LAA data augmentation method with the advanced ViT-Triplet architecture, incorporating the ViT model and hard triplet loss. This approach achieves impressive results with CMC top-1, CMC top-5, mAP top-5, and mAP top-10 scores of 96.11%, 99.86%, 96.42%, and 96.01%, respectively. On this basis, the algorithm introduces text background removal during the inference stage, further improving overall performance with accuracy, precision, recall, and F1 score of 98.47%, 98.61%, 98.33%, and 98.47%, respectively.