Open set semantic egmentation of remote sensing images

• Autor(a) principal: Caio Cesar Viana da Silva
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/33529
• topic: Open Set; Deep Learning; Semantic Segmentation; Remote Sensing; Computação - Teses; Visão por computador; Aprendizado do computador; Sensoriamento remoto; Linguagem de programação (Computadores) - Semântica

The main approaches developed in computer vision and digital image processing are focused on data obtained through smartphones and compact cameras. These cameras are typically used to capture scenes on RGB channels, only in the visible spectrum. Another source of images that are exploited by computer vision is satellite images or aerial images. However, the development of computational vision approaches that exploit satellite imagery is relatively recent, mainly due to the limited availability of this type of image. Until recently they were exclusively for military use. Access to aerial imagery, including spectral information, has been increasing mainly due to the low cost of drones, new civilian satellites, and data sets on various public platforms. In the area of remote sensing, applications that employ computational vision techniques are modeled for classification in closed set scenarios. However, the world is not purely closed set, many scenarios present classes that are not previously known by the algorithm, an open set scenario. Thus, the main objective of this dissertation is the study and development of semantic segmentation techniques considering the open set scenario applied to remote sensing images. The main contributions of this dissertation are: (1) a discussion of related works, showing evidence that semantic segmentation techniques can be adapted for open set scenarios; (2) the development of two methods for open set semantic segmentation. The OpenPixel and OpenFCN methods presented competitive results when compared to the closed set methods in the same data set. On average, the OpenPixel method had an overall accuracy of 57.51\%, a normalized accuracy of 54.23\% and a Kappa Index of 0.5602. For OpenFCN, the method resulted in an overall accuracy of 82.27\%, a standard accuracy of 64.39\% and a Kappa Index of 0.7630. It is possible to conclude that the proposed methods can segment unknown classes while still correctly classifying most of the known classes, performing open set semantic segmentation on remote sensing images.