Face recognition by means of advanced contributions in machine learning

Resumen

Face recognition (FR) has been extensively studied, due to both scientific fundamental challenges and current and potential applications where human identification is needed. FR systems have the benefits of their non intrusiveness, low cost of equipments and no useragreement requirements when doing acquisition, among the most important ones. Nevertheless, despite the progress made in last years and the different solutions proposed, FR performance is not yet satisfactory when more demanding conditions are required (different viewpoints, blocked effects, illumination changes, strong lighting states, etc). Particularly, the effect of such non-controlled lighting conditions on face images leads to one of the strongest distortions in facial appearance. This dissertation addresses the problem of FR when dealing with less constrained illumination situations. In order to approach the problem, a new multi-session and multi-spectral face database has been acquired in visible, Near-infrared (NIR) and Thermal infrared (TIR) spectra, under different lighting conditions. A theoretical analysis using information theory to demonstrate the complementarities between different spectral bands have been firstly carried out. The optimal exploitation of the information provided by the set of multispectral images has been subsequently addressed by using multimodal matching score fusion techniques that efficiently synthesize complementary meaningful information among different spectra. Due to peculiarities in thermal images, a specific face segmentation algorithm has been required and developed. In the final proposed system, the Discrete Cosine Transform as dimensionality reduction tool and a fractional distance for matching were used, so that the cost in processing time and memory was significantly reduced. Prior to this classification task, a selection of the relevant frequency bands is proposed in order to optimize the overall system, based on identifying and maximizing independence relations by means of discriminability criteria. The system has been extensively evaluated on the multispectral face database specifically performed for our purpose. On this regard, a new visualization procedure has been suggested in order to combine different bands for establishing valid comparisons and giving statistical information about the significance of the results. This experimental framework has more easily enabled the improvement of robustness against training and testing illumination mismatch. Additionally, focusing problem in thermal spectrum has been also addressed, firstly, for the more general case of the thermal images (or thermograms), and then for the case of facialthermograms from both theoretical and practical point of view. In order to analyze the quality of such facial thermograms degraded by blurring, an appropriate algorithm has been successfully developed. Experimental results strongly support the proposed multispectral facial image fusion, achieving very high performance in several conditions. These results represent a new advance in providing a robust matching across changes in illumination, further inspiring highly accurate FR approaches in practical scenarios.