GENHAND: A HYBRID EVOLUTIONARY ADVERSARIAL FRAMEWORK FOR STRUCTURE-PRESERVING CHARACTER LEVEL GENERATION

Abstract

The Focus of the proposed research is on the generation of Handwritten Text, which has been a difficult problem because of style variations, overlapping strokes and the problem of preserving structural integrity at the character level. This research work focusses on a hybrid evolutionary-adversarial framework that can be used to solve such issues by integrating object detection, evolutionary optimisation and generative modelling. To train a YOLOv8 detector was created a custom dataset of 28,830 annotated English alphanumeric characters, and Particle Swarm Optimisation (PSO) and Covariance Matrix Adaptation Evolution Strategy (CMA-ES) are applied to perform the refinement of the bounding boxes and correct their orientations. To produce high-quality handwriting, a two-GAN architecture, ContourGAN to generate edge fidelity and FullMaskGAN for structural completeness that is jointly stabilized with genetic algorithm is proposed in this work. A Lightweight Transformer Module is developed to processes the output after they have been refined to enhance the perceptual quality. The resulting pipeline allows end-to-end conversion of typed text to a variety of, structure-preserving handwriting styles, with several applications in data augmentation, OCR enhancement, and digital archiving.