Fast eye localization from thermal images using neural networks

Abstract

The paper presents a fast algorithm for eye localization from thermal images. Due to blurred edges, lower quality of thermal images compared with visible-light images, the process may be more complicated. On the other hand, it seems relatively easy to designate specific areas of the eyes because of their relatively highest brightness resulting from the highest temperature in these face areas. It turns out, however, that the highest pixel brightness does not always determine unequivocally the position of the eyes. Therefore, the algorithm described here proposes a specialized method for segmentation of images containing eye areas and a specialized classifier. The first part of the algorithm is a block for generating features describing eye areas. The second stage is a decision block built based on a neural network allowing for effective classification of pre-designated areas. A suitable combination of these blocks into a single system allowed for correct analysis of images with rich geometric diversity (scale, position, orientation) and brightness distribution, reaching more than 91 % of correct localizations and time of analysis of a single image of hundredths of a second. Algorithms of this type can be used as the initial stages of facial recognition, emotion recognition, quantitative analysis of the face or modern human-computer interfaces. The described algorithm may be used for both conventional thermal images and thermograms (after prior conversion of temperature values into brightness values), which affects its potential application in medicine. The benefits of this study are twofold: fast localization of characteristic points of the eyes and the possibility of processing despite the typical problems of thermovision and image transformations.