Object detection and extraction from fixed-layout documents using Deep Learning
Technical or scientific documents usually have a lot of figures, tables or diagrams besides text. In one of our systems it was crucial to extract these kind of objects from a PDF document as a separate image and take note of its location in the original document.
We have implemented the YOLO detection algorithm and evaluated multiple architectures of the underlying convolutional neural network.
We have pretrained the model on a large set of labeled images and then applied transfer learning to our specific task.
To increase the data set for our specific task, we have applied data augmentation.
Using Deep Learning mechanism we created system which can detect and extract objects from fixed-layout documents.