Deep Multipath Network Architecture for Recognizing Visual Objects in Document Images

Abstract

Automated document processing systems are crucial tools for processing large volumes of document images. The primary visual elements present in these documents include stamps, logos, and signatures, which serve various purposes. This research proposes a novel multipath architectural network for simultaneous and accurate recognition of visual objects. This architecture uses deep learning and image patches as core components through three basic paths. The image characteristics extracted from these paths are combined to achieve the final feature vector, which is subsequently processed to determine the class of each image patch. The proposed architecture is validated through experiments conducted across four scenarios using the SPODS and Tobacco 800 datasets. The comprehensive evaluations demonstrate remarkable outcomes, achieving precision, recall, and accuracy rates of (99.54%). Furthermore, the results demonstrate that visual object recognition tasks, including the identification of signatures (100%), stamps (99.10%), and logos (99.30%), are executed with exceptional accuracy. In addition, the non-visual object recognition of text objects achieves a perfect score of (100%). To sum up, the findings of these evaluations showed significant improvements in accuracy compared with competitors.