shraddhaA2/handwritten-amharic-character-recognition

Deep learning-based recognition of handwritten Amharic characters using image classification. # Handwritten Amharic Character Recognition Through Transfer Learning **Integrating CNN Models (ResNet50) and Machine Learning Classifiers (SVM)** --- ## 1. Project Title & Objective ### Title **Handwritten Amharic Character Recognition Through Transfer Learning: Integrating Pretrained CNN Feature Extraction and Support Vector Machine Classifiers** ### Objective To build an accurate, reproducible, and computationally efficient machine learning pipeline for recognizing handwritten Amharic characters across all **238 character classes** (34 base consonants × 7 vowel orders). The system leverages **Transfer Learning** by using a deep ImageNet-pretrained **ResNet50** convolutional neural network to extract 2,048-dimensional feature representations from preprocessed handwritten strokes, which are subsequently classified using a **Support Vector Machine (SVM)** with an RBF kernel. --- ## 2. Dataset Description & Statistics The dataset is obtained from the standard benchmark: Fetulhak/Handwritten-Amharic-character-Dataset - **Total Images**: **37,804 images** - Dataset 1: 18,902 images - Dataset 2: 18,902 images - **Number of Categories**: **238 unique classes** (34 base consonant families × 7 vowel orders) - **Image Resolution**: 28 × 28 pixels - **Color Mode**: Grayscale / Luminance - **Format**: JPEG - **Data Collection Details**: Cropped from scanned handwriting forms contributed by individuals across diverse age ranges, education levels, and both right- and left-handed writers. - **Filename Convention**: ` . .jpg` (e.g. `001he.1.jpg`, `001he.137.jpg`, `238po.93.jpg`). --- ## 3. System Architecture & Methodology The machine learning workflow follows a multi-stage transfer learning pipeline: ``` [ Handwritten Character Image (28×28 Grayscale) ] ↓ [ Preprocessing Pipeline ] 1. Grayscale conversion & noise reduction (Gaussian blur) 2. Adaptive Otsu binarization 3. Stroke contour bounding-box detection & aspect-ratio padding 4. Resizing to 224 × 224 pixels & 3 …