# ✍️ Handwritten Signature Classifier ### AI-powered handwritten signature recognition for document processing An image-classification model designed to distinguish handwritten signatures from non-signature content in scanned documents and document images. Built on the **MobileNetV3-Small** architecture, this model is intended for efficient inference and integration into automated document-processing workflows. --- ## 🚀 Overview The model classifies an input image into one of two categories: | Class | Description | |---|---| | `SIGNATURE` | Handwritten signature or signature-like handwriting | | `NOT_SIGNATURE` | Non-signature document content, such as printed text, dates, numbers, and other visual elements | Its primary purpose is to help document-processing systems identify genuine signature regions and reduce false-positive detections. ## 🧠 Model Architecture - **Architecture:** MobileNetV3-Small - **Base Model:** `qualcomm/MobileNet-v3-Small` - **Task:** Binary image classification - **Framework:** Transformers - **Input:** RGB image - **Output:** Class probabilities and predicted label - **Classes:** SIGNATURE / NOT_SIGNATURE MobileNetV3-Small provides a lightweight architecture suitable for resource-constrained environments and desktop inference. ## 📚 Training Dataset **Dataset:** [`AmjadKha/HandWrittenSigntureDataset`](https://huggingface.co/datasets/AmjadKha/HandWrittenSigntureDataset) The model uses a handwritten signature dataset for training and evaluation. Supported document languages: - English - Arabic ## 📊 Evaluation Metrics The model is evaluated using the following metrics: | Metric | Purpose | |---|---| | Accuracy | Measures overall classification correctness | | Precision | Measures how many predicted signatures are actually signatures | | Recall | Measures how many actual signatures are detected | | F1-score | Balances precision and recall | > Actual evaluation scores should be added after testing on a separate held-out test dataset. ## ⚙️ Installation ```bash pip install transformers torch pillow ``` ## 💻 Usage ```python from transformers import pipeline classifier = pipeline( "image-classification", model="AmjadKha/YOUR-MODEL-REPOSITORY" ) results = classifier("signature.png") for result in results: print(result["label"], result["score"]) ``` Replace `YOUR-MODEL-REPOSITORY` with the actual Hugging Face model repository name. ## 🏦 Intended Applications - Banking document processing - Signature verification workflows - Account-opening documentation - Customer onboarding documents - Automated document classification - PDF signature-region filtering - Intelligent document processing (IDP) ## ⚠️ Limitations - Classification does not establish the legal authenticity of a signature. - Handwritten names and initials may resemble signatures. - Performance depends on image quality, resolution, and training-data diversity. - The model should be evaluated on representative documents before production deployment. - It should be combined with signature detection and verification mechanisms for high-assurance workflows. ## 📄 License Apache License 2.0. ## 👨‍💻 Author **Amjad Khaliliah** Hugging Face: [@AmjadKha](https://huggingface.co/AmjadKha) --- *Lightweight computer vision for intelligent document processing.*