Triplet CNN-based word spotting of Gujarati Handwritten Document

Word spotting of Gujarati handwritten documents is a highly challenging task due to the complexity of the handwritten text in the Gujarati language. This paper presents a novel approach to word spotting, which includes a transfer learning-based triplet CNN model for generating word embeddings and calculating triplet loss. During training, the CNN processes a set of triplet samples consisting of anchor, positive, and negative images. Triplet loss is then calculated to create word embeddings by minimizing intra-class distances and maximizing inter-class distances, which are further evaluated using Euclidean distance for comparison. The proposed approach is tested on handwritten image documents, and the experimental results demonstrate its effectiveness.