Detecção de placas veiculares em ambientes aéreos baseado em métodos de aprendizagem profunda

Abstract

In this dissertation, considering the difficulties of detecting small objects due to the low visual quality and spatial resolution of the images, as well as the little context information, we propose a methodology for detecting license plates in aerial environments, where the object of interest has low pixel resolution compared to the input image. The proposed methodology is composed of three distinct systems. Two use image quality improvement techniques and one does not. The main contribution of this dissertation is the organization of a database composed of aerial images to evaluate the performance of the proposed methodology. As a second contribution, there is the use of visual quality improvement methods based on learning and digital image processing to detect license plates in aerial environments. As obtained results, it appears that the proposed systems present similar results, in terms of global accuracy. When analyzing the results by group of images, it is observed that the proposed system with histogram equalization presents the best results with an accuracy of up to 85.67% in adverse sunny conditions and 99.33% in adverse shaded conditions.