Author: Antoni Rodríguez Villegas
Virtual Room: https://eu.bbcollab.com/guest/6c3cff57ecf0466eabc44e6df8ce7cbc
Date & time: 22/09/2021 – 10:15 h
Session name: Text Recognition
Supervisors: Marçal Rossinyol, Dimosthenis Karatzas
Abstract:
Automatic Number Plate Recognition (ANPR) systems are an active research topic due to their many applications and the present availability of data and high computational power.Current ANPR systems consist of two main steps: detection and recognition. The quality of the detections affect greatly that of the recognition, since the image is cropped based on what is believed to be the area of interest. Also, these systems tend to be computationally exhaustive, forcing them to be deployed in dedicated and expensive hardware.In this work, we aim to develop an efficient and lightweight end-to-end trainable model that is able to detect and read a license plate in a single step, eliminating detection-recognition dependencies. For this purpose, we created a neural network architecture with independent branches for each task that can be trained end-to-end. Finally, we quantised and tested this model in order to be deployed in on-the-edge devices.
Committee:
– President: David Masip(UOC)
– Secretary: Angel Sappa(UAB)
– Vocal: Ramon Morros(UPC)