Image Enhancement using GAN-based Augmentation in CNN Models for Classification Dyslexia Severity

Dyslexia, a complex neurodevelopmental disorder, poses formidable challenges to individuals in acquiring proficient reading skills despite adequate instruction and cognitive abilities, with its accurate diagnosis remaining elusive due to its multifaceted nature and symptom overlap with other learning disorders. This study endeavours to develop an automated dyslexia diagnosis system leveraging dyslexic handwriting images and cutting-edge deep-learning methodologies, focusing on implementing advanced techniques, notably Convolutional Neural Networks (CNNs), to construct a robust dyslexia diagnosis model. To counteract data imbalance challenges inherent in dyslexic handwriting datasets, data augmentation utilizing Generative Adversarial Networks (GANs) will be integrated to enhance model accuracy and generalization capability. The project's outcomes encompass a meticulously trained CNN-based model adept at accurately categorizing dyslexia severity levels. Experimental findings demonstrate the successful implementation of automated dyslexia severity classification using advanced deep learning techniques and GAN-based data augmentation, wherein the LeNet-5 model, trained with an augmented dataset enriched with additional words generated by GANs, outperformed both the original DenseNet201 model and the LeNet-5 model trained solely with original and generated words within the original dataset's existing classes without introducing new words, achieving an impressive F1 score, recall, precision, and accuracy of 91 %, underscoring the efficacy of data augmentation techniques in enhancing the accuracy and robustness of CNN models for dyslexia diagnosis and classification.