MFFCNN: Multi-Feature Fusion-based Deep Learning Architecture for Typewritten Word Recognition System

For the preservation of cultural legacy, to provide accessibility, to support language learning, and to facilitate effective data processing, accurate word recognition in Gurmukhi, particularly for typewritten text, is of utmost importance. The precise identification of diacritics, dealing with problems like broken or fading letters, and adjusting differences in ink density are a few difficulties this recognition method faces. For Gurmukhi language writings to be seamlessly digitized, translated, and distributed and, in turn, for more people to have access to and comprehend this rich linguistic legacy, these challenges must be overcome. This work contributed to the same and proposed a multi-feature fusion-based deep learning architecture (MFFCNN) that extracts both handcrafted and CNN-based features, then fuses all the features and uses them further for classification and recognition. The dataset of Gurmukhi typewritten content is captured using a mobile camera from different dissertation files, books, novels, etc., available in typewritten. The dataset is pre-processed, and segmentation is performed before MFFCNN for word recognition and extraction. The performance of the proposed method is computed based on the accuracy, and a comparison is made with the CNN-based model. The results of the proposed method presented the effectiveness of multi-feature fusion with deep learning architecture specifically for text recognition.