Certain investigation on non cursive and cursive handwritten character recognition using deep learning models

Abstract

Optical Character Recognition (OCR) is a critical area in computer vision, enabling the transformation of text from images into editable and searchable formats, with applications ranging from banking and information retrieval to vehicle number plate recognition. This work focuses on two major challenges in OCR: Non-Cursive Character Recognition (NCCR) and Cursive Handwriting Recognition (CHR), each requiring specialized strategies due to the distinct characteristics of the input text. For NCCR, the primary challenge lies in processing varied formats that may include different orientations, languages, characters, numbers, and mathematical symbols. To address this, a Gaussian Amended Bilateral Filter (GABF) is proposed for preprocessing, effectively reducing noise while preserving essential edges. For feature extraction, an Enhanced Haar Wavelet Transform (EHWT) captures discriminative structural, textural, and edge-based features representing both global and local character patterns. The recognition stage employs a novel Attentional Coati Separable Dense Convolutional Network (ACSDCN), which combines attention mechanisms, separable convolutions, and dense connections to enhance recognition accuracy while maintaining computational efficiency newline