EXPLORING THE INFLUENCE OF LEARNING RATE ON THE DEEP LEARNING CLASSIFICATION PERFORMANCE: A TIFINAGH HANDWRITTEN CHARACTERS CASE STUDY

Abstract

Pattern recognition remains an essential field in the world of scientific research, particularly with the development of new technologies such as machine learning and deep learning, and their application in various aspects of life. Indeed, the recognition and classification of handwritten characters have earned considerable attention in researchers' studies. In this context, we present a comparative study of the most widely used deep learning convolutional neural networks' (CNNs) architectures, including DenseNet201, Inception_Resnet_V2, Inception_V3, MobileNet_V2, ResNet50, VGG16, and VGG19, to automatically recognize and classify Tifinagh handwritten characters. The proposed paper has been tested using the Amazigh Handwritten Character Data-base (AMHCD). This work examines the impact of learning rate on classification performance and analyzes the added value of using a cyclical learning rate. A statistical study, in particular the Scott-Knott algorithm and Borda Count method, applied to the obtained results, displays that the fine-tuned version of Inception_Resnet_V2 with a learning rate of 0.0001 and the use of cyclical learning rate yields the best classification performance. This performance is shown by an accuracy of 99.66%, sensitivity of 99.95%, specificity of 99.99%, precision of 99.64%, and an F1 score of 99.65%.