Análise estatística das taxas de reconhecimento de imagens de edifícios em ambientes urbanos utilizando quantização vectorial e aprendizagem profunda

Abstract

Building recognition is essential for a variety of applications such as automatic target detection, 3D city reconstruction, digital navigation, etc. This paper aims to comparatively analyze the recognition rates of building images, using the Vector Quantization technique for image compression using the Linde-Buzo-Gray algorithm, with the results obtained by the Deep Learning method. Forty classes were analyzed, with 30 images per class, separately, in gray, red, green, blue and RGB, varying the number of centroids in 16, 32, 64, 128 and 256 for the vector quantization technique, and also varying the percentage of the amount of images for training in 40%, 50% and 60%, with their respective percentages of the amount of images for recognition, in both methods. To check the differences, ANOVA was performed, with Tukey's post-hoc at 5% significance. The descriptive results showed high recognition rates, in both methods. In the inferential analysis of the results obtained in Vector Quantization, significant recognition rates were found from 32 centroids on. When comparing the results of the two techniques, no significant difference was found.