Grad-Cam Visualization Of Arabic Letter Character Prediction

Arabic letters, commonly called hijaiyah letters, present a considerable challenge in acquisition and mastery. Introducing hijaiyah letters is a significant subject due to the inherent challenges associated with their composition. This study aims to compare the class activation visualization in Arabic characters by employing a custom model and contrasting it with the widely used models, namely AlexNet and LeNet. The employed model utilizes the Class Activation Mapping (CAM) visualization technique to demonstrate its understanding of the character identification process effectively. This approach facilitates the observation of key focal points when the model identifies a certain character. This study aims to identify key elements that contribute to the effectiveness of a Convolutional Neural Network (CNN) model in accurately recognizing characters. This will be achieved by training the CNN model using a substantial dataset that specifically emphasizes Arabic character recognition. The study will employ Class Activation Mapping (CAM) to visualize the results. The results of this study will not only offer a comprehensive comprehension of the model’s approach to Arabic character detection. However, they will also assist in identifying any problems that may arise during this procedure. The outcomes of this research would enhance the model’s capacity to accurately identify Arabic script, hence facilitating the implementation of assistance for handling Arabic text that is damaged or blurred. In this investigation, it was observed that the performance of the custom model surpassed that of the AlexNet and LeNet convolutional neural network (CNN) models. Training on a dataset consisting of 13,440 data points, the custom model achieved a notable accuracy rate of 97.38%. Additionally, the model exhibited a loss of 9.07% at epoch 50. In the interim, AlexNet and LeNet demonstrated accuracy of 96.15% and 93.12%, with losses of 15.88% and 21.90%.