As large amounts of technical documents have been published in recent years, efficiently retrieving relevant documents and identifying locations of targeted terms are urgently needed.
We propose an end to end system that takes a picture of printed document and outputs the locations of math formulas in the image. In our project we use a pretrained model Yolov5 as feature extractor and detector.
Data is collected from ICDAR competition for the 2 years 2019 and 2021. Data can be found here: https://www.kaggle.com/ro101010/math-formula-detection
We managed to achieve a precision of 0.943 , a recall of 0.912 and mean average precision @ IOU = 0.5 of 0.949.
Clone the project repository to your machine In your terminal run the following commands:
- cd cloned-project-directory/
- pip install -r requirements.txt
- streamlit run app.py
A link to the app will appear in the terminal click on it : http://localhost:8501/
You will be directed to the main app page:
Click Launch Detection,the output should look like this
You can also choose to upload a pdf and repeat the same steps.