CNN comparisons models on dyslexia handwriting classification / Iza Sazanita Isa ... [et al.]

Dyslexia is developed by neurobiological in origin which is categorized as learning disorder that affect the ability to read, spell, write and speak. The most common dyslexia symptom can easily be identified through the handwriting pattern. There are many intelligence and computational methods that have been proposed, and they have provided various and different performance to evaluate the proposed system ability. However, system performances are varied and nonstandardized in each assesment on dyslexic children to validate the presence of dyslexia symptom. The recent deep learning models have been employed to improve the assesment performance and (the models/ they have shown) shows significant output to detect and classify the present of dyslexia symptoms among school children. Therefore, there is a crucial need in deep learning, specifically for Convolutional Neural Network ( CNN) to validate performances of different networks, so that the most performed CNN could be a bench mark in evaluation to detect such symptom. This study aims to compare different deep learning networks specifically the CNN models to validate its performance in terms of the capability to classify dyslexic handwriting among school children. This study is proposed to compare different CNN models such as CNN-1, CNN-2, CNN-3 and LeNet-5. The proposed methods to compare the CNN performances are developed by using Jupyter notebook as platform. Meanwhile, keras is the higher-level API framework to provide a more flexible way for defining models. It specifically allows to define multiple input or output models as well as models that share layers. The tensorflow is also used for machine learning applications such as neural networks. Before that, the dataset of handwriting image is preprocessed by the augmentation process which includes the rotation of all images. CNN models have shown significant performance and provided sufficient results of performance with more than 87% of accuracy in classifying the potential dyslexia symptom based on handwritten images