PROJECT TITLE :

Sensitivity Analysis of Influence Quantities on Signal-to-Noise Ratio in Face-Based Recognition Systems - 2017

ABSTRACT:

Nowadays, face recognition systems are going to widespread in many fields of application, from automatic user login for financial activities and access to restricted areas, to surveillance for improving security in airports and railway stations, to cite a few. In such situations, several architectures based mostly on both 2D image analysis and 3D reconstruction are investigated and proposed in literature. The actual performance of such systems in terms of correct call rate is full of many quantities of influence mainly concerning the conditions of acquisition of the image to be processed. As an example, the image luminosity, the lens defocus and also the movement of a subject throughout the image acquisition will be sources of uncertainty that propagate up to the ultimate classification result, therefore affecting the reliability of a topic identification. In previous papers, the authors proposed appropriate uncertainty models for each 2D and 3D based mostly architectures in a position to quantify on-line the extent of confidence to assign to the output of such systems in keeping with the ISO-GUM. The proposed models required, for every amount of influence, to estimate separately their deviations with respect to the reference values achieved in ideal acquisition conditions during the training section. On the other hand, the quality of a picture may be linked to the a lot of general concept of signal-to-noise ratio (SNR), as a result of noise affects the pixel of the image, therefore introducing uncertainty on the final image. So, wanting for the development of a a lot of straight and simple to use uncertainty model, in this paper the relationships among the quantities of influence and the image SNR are investigated. This activity represents the primary step toward the conclusion of face-based recognition systems able to assign a level of confidence to the output results starting solely from the evaluation of SNR on the input image.