Handwritten Amharic Character Recognition Through Transfer Learning: Integrating CNN Models and Machine Learning Classifiers

Abstract

Handwritten Amharic character recognition presents significant challenges due to the script’s syllabic nature and variations in handwriting styles. This study investigates a hybrid approach that integrates convolutional neural networks (CNNs) with machine learning classifiers to enhance recognition accuracy. Transfer learning is applied using four CNN architectures: AlexNet, VGG16, VGG19, and ResNet50 as feature extractors. Initially, their performance is evaluated with softmax classifiers. Subsequently, the softmax layer is replaced with machine learning classifiers, including Random Forest, XGBoost, and Support Vector Machine (SVM), while freezing the pretrained feature extractors. The Hybrid ResNet50 + SVM model achieves the highest accuracy of 91.89%, with a precision of 92.46%, recall of 91.15%, and an F1-score of 91.80%. These results indicate that SVM serves as a potential alternative to softmax, offering robust classification performance for complex handwritten scripts. This research contributes to advancements in handwritten character recognition systems for underrepresented languages.