Abstract:
In this document, we present an object detection system for italian documents. Unlike similar systems which use a single deep-learning solution, this system employs different solutions
for a fast and accurated detection. The first is an image segmentation module which process an acquired-scanner image and find every important artificats. The second
is a custom CNN for detect every artificat's rotation and then use the information for set the document to the upright (this is important for read the document-text content).
The third is a simple CNN for detect each upright element. We present the algorithm used in the first part and the training methods for both types of networks.
We also perform analysis on the networks, and present empirical results on a large test set. Finally, we present preliminary results for detecting documents.
