During the implementation of our final year project, we focused on automating document classification to meet the needs of the company EdenAI. Our main objective was to create an API capable of taking a document as input and returning the corresponding type among the following categories : handwritten, id, invoice, receipt, resume, or table. To achieve this goal, we explored different solutions and approaches, relying on the latest technological advancements in the field, such as automatic document understanding.
We considered several approaches for our project, starting with a direct approach using the ResNet model for image classification. Then, we developed a more customized solution by implementing a classification pipeline. This pipeline combines two models : a vision model and a natural language processing (NLP) model. Lastly, we explored an end-to-end approach by training the Donut model. This model does not rely on OCR but uses a vision encoder to process the image and an NLP decoder to process the extracted text.
We then evaluated our different solutions using classification metrics such as recall, precision, accuracy, specificity, and F1 score. Due to its ability to generalize across different languages, the Attention ResNet model proved to be the best solution for deployment, offering optimal performance and efficient execution time.
Key words : Deep Learning, Natural Language Processing (NLP), computer vision (CV), Document understanding, OCR.
Use these commands to install the dependencies on a virtual environment.
python3 -m venv eden_envpip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpupip3 install -r requirements_linux.txtdocker build -t user_name/edenai_app .docker run -d -p 8000:8000 image_nameRun the file test.py to try our API. Make sure to change the ip address of the endpoint link if necessary.
