Deep Embedding Learning for Printed Word Spotting using Triplet CNNs and Transfer Learning

Word spotting plays a crucial role in the broader domains of Optical Character Recognition (OCR) and Natural Language Processing (NLP), particularly when dealing with low-resource languages like Gujarati. The inherent complexity of Gujarati—stemming from its rich morphological structure and intricate character forms—presents significant challenges for automated recognition systems. This study investigates the application of Triplet Convolutional Neural Networks (CNNs) for learning discriminative feature embedding tailored to the task of Gujarati word spotting. By employing a triplet loss function, the proposed approach aims to minimize intra-class variation while maximizing inter-class separation within the learned embedding space. To rigorously evaluate the effectiveness of the method, experiments were conducted on both printed and handwritten Gujarati word datasets. The framework incorporates established deep CNN architectures such as ResNet50, VGG19, and InceptionV3 within the Triplet CNN configuration. Furthermore, a comparative analysis of different similarity metrics—including Euclidean, Cosine, and Manhattan distances—was performed to determine the most suitable measure for evaluating embedding similarity. The results demonstrate that the Triplet CNN model, particularly when paired with Euclidean distance, consistently outperforms baseline architectures and alternative distance metrics. These findings underscore the potential of the proposed approach in improving accuracy and reliability for Gujarati word spotting across diverse document formats.