Bengali License Plate Recognition from Noisy Video Footage Using Deep Learning

Abstract

At present, the license plate recognition of a vehicle from noisy footage has become an important issue in Bangladesh. Nowadays, in Bangladesh, the number of vehicles is increasing at a very high frequency. Moreover, unlike a US license plate, a Bengali license plate consists of two lines of inputs. As the vehicle number is increasing the difficulty of identifying a vehicle is also increasing. There are multiple problems where number plate recognition is a real necessity such as, in a crime scene, finding out a lost vehicle, identifying a guilty vehicle in a road accident, etc. The main challenge in this system is to predict the Bengali numbers from a very noisy image. Most of the methods for identifying vehicles from noisy data are not as accurate as they should be. In order to reduce such noise, this research explores different filtration algorithms which are edge-preserving. The research also gives a modified version of the non-local mean denoising filter which provides a significant amount of good results in terms of detecting Bengali words, characters, and digits from Bengali license plates using YOLO Version 3 algorithm.