Reconhecimento de placas veiculares em cenários complexos utilizando o método do subespaço

Abstract

In this dissertation, a complete system is proposed for carrying out the process of detection and recognition of vehicle license plate in images where the acquisition process was carried out with the camera and vehicle in motion, and which has variations in lighting and resolution, as well as complex scenarios. As pre-processing steps, conversion to gray scale and YOLO object detector were used to perform car detection. The first stage of the proposed methodology was the creation of image bases, in which the YOLO object detector was used to detect the license plate and six types of degradation for data augmentation: Gaussian noise, Poisson noise, laplacian noise, scale change, rotation and contrast change. Then the proposed methodology consisted of the application of two techniques for vehicle license plate recognition, in the first technique the mutual subspace method is used, and in the second method convolutional neural networks are used as baseline. The results obtained in the approach using the mutual subspace method had as the best result the accuracy of 57% and average prediction time of 0.33ms, while the result of the approach based on convolutional neural networks obtained as the best result the accuracy of 94% and average prediction time of 200ms.