Segmentation-Based Word Spotting in Handwritten Ayurveda Manuscripts Using Siamese Convolutional Network

Word spotting is typically formulated as an image-based retrieval problem, wherein one or more query word images are systematically compared against a repository of candidate word images to identify instances exhibiting visual or semantic similarity. This task follows the query-by-example paradigm, where the goal is to retrieve instances from the database that match the query based on similarity. This paper presents a segmentation-based word spotting framework specifically designed for Ayurveda manuscripts, which frequently exhibit ambiguous or indistinct word boundaries. A new dataset tailored for word spotting is also introduced. The system spots words using proposed Siamese convolutional network. Our experimental results, evaluated on a complex multi-writer corpus, demonstrate that the proposed framework consistently outperforms traditional matching approaches. The system achieved promising results in word spotting for handwritten Ayurveda manuscripts when tested on this new dataset.