An efficient word-spotting model for Gujarati language using optimization driven deep learning technique

Resumen

Digitization of historical handwritten documents is highly preferable as it provides access to a wide range of documents to huge users. But, the information retrieval from such documents is a tough task as it requires full-text search. Word spotting is a process or retrieving the content from such documents based on a query image. This process results in a ranked list of words that are similar to the query word image. In several existing works, the discrimination between the feature vectors are not well understood. To resolve this issue, this paper deals with word spotting of Gujarati language. Initially the text document is pre-processed based on filtering, binarization and skew detection. Then, the pre-processed image is subjected to segmentation where line and word segmentation are performed to segment the words separately from the image. Then, the feature vectors are extracted from the word images using the proposed 3-layer deep convolutional autoencoder (CAE) and the model is further tuned using the improved elephant herding optimization (IEHO) algorithm. The feature vectors are then indexed using the geometric hashing algorithm and finally the matching is carried out based on Euclidean distance, Minkowski distance, hash similarity and Manhattan distance similarity measures. The framework is evaluated and compared with the existing models and the results proved the efficacy of the proposed model over the other existing models.