Detecção de vazamento de gás natural em imagens usando filtro novidade

Abstract

This dissertation proposes a method of detecting natural gas leak in oil installations, specifically in onshore oil wells, applying pattern recognition techniques in digital images. Other methods of detecting natural gas leak found in the literature, has as its foundation the analysis of methane molecules, applying the catalytic principle or through the absorption of infrared spectrum. These methods have some disadvantages such as reduced service life and conditions of false negatives, respectively. The method of detecting natural gas leak proposed is based on the application of the concept of novelty filter in digital images originated from CCTV. For this, it was tested six different types of novelty filters. The difference between these filters is in the component type of color spaces (RGB or HIS) used in forming the training base of the respective filter. The best component that characterizes a natural gas leak was determined by performance measuring using ROC curves. The results obtained in the filter of novelty R are promising, with maximum sensitivity and specificity equal to 0,969 and 0,961 respectively, and average AUC of 98.527%.