natenaile/Handwritten-Amharic-Character-Recognition

Combining CNN-based feature extraction with classical machine learning classifiers enables accurate and robust recognition of handwritten Amharic characters. # Handwritten Amharic Character Recognition through Transfer Learning: Integrating CNN Models and Machine Learning Classifiers. ## Overview This project focuses on automating the recognition of handwritten **Amharic characters**, which are a part of the Ethiopian script. The Amharic script consists of 238 unique characters, including 34 basic characters with seven variations representing different vowel sounds. Due to the syllabic nature of the script and variations in handwriting styles, recognizing handwritten Amharic characters is a challenging task. In this study, we experiment with four convolutional neural network (CNN) architectures—AlexNet, VGG16, VGG19, and ResNet50—utilizing transfer learning for the recognition of handwritten Amharic characters. Initially, we evaluate the performance of the default softmax classifiers in these models. Subsequently, we replace the softmax classifiers with machine learning classifiers, specifically Random Forest, XGBoost, and Support Vector Machine, to compare their performance. The system achieves accuracy, precision, recall, and F1-score values of 91.89%, 92.46%, 91.89%, and 91.80%, respectively. ## Methodology 1. **Dataset Overview**: - **Total Samples**: 37,752 - **Training Samples**: 30,201 - **Validation Samples**: 7,551 - **Image Size**: 28x28 pixels - **Image Type**: JPEG 2. **Data Augmentation**: To enhance the robustness and generalization of the models, several **data augmentation** techniques are employed on-the-fly during training, rather than pre-saving new images. These techniques include: - **Random Rotation**: The images are randomly rotated within an angle range of ±10 degrees to introduce variability in orientation. - **Affine Transformations**: Techniques such as translation (shifting) by 10% of the image dimensions are applied to create slight variations in image positioning. - **Random Resized Cropping**: This involves cropping the images from random locations while ensuring that the main part of …