Face recognition in a wild

Abstract

This thesis deals with face recognition in uncontrolled environments. newlineUncontrolled environments contain more variation in pose, lighting, expression, newlineocclusion, background, image quality, scale, and makeup and hence face recognition in newlineuncontrolled environments is much more challenging than in controlled conditions. In newlineaddition, many real-world applications require good recognition performance in newlineuncontrolled environments. Example applications includes automotive drowsiness newlinedetection system, AI based car theme personalization, airport surveillance and newlinemonitoring, premise/office control systems governments initiatives like Aadhaar card, newlineproperty registration/validation and social networking applications. Therefore, researcher newlinescholars and innovative niche startup companies have shifted their interest from newlinecontrolled environments to uncontrolled environments over the past one-decade. newlineAfter doing extensive literature survey, we find that face recognition in unconstrained newlineenvironments is still an unsolved problem although many face recognition algorithms newlinehave been proposed which tried to address problems in a limited manner. Many existing newlinealgorithms are mainly having three limitations. newline1. Algorithms do not perform well in uncontrolled environments and recognition rate newlineis very poor. newline2. Algorithms cannot handle environments parameters in totality. newline3. Algorithms are very slow and hence not suitable for real time applications. newlineIn this thesis, we investigate various algorithms and various associated challenges for newlineface recognition. To improve the face recognition rate in an uncontrolled environment, we newlinehave proposed algorithm based on combined Haar features (based on 6 rectangle newlinewindow). It also helps in improving overall speed of executions due to optimized newlineoperations and search. Proposed algorithm is integrated with illumination normalization newlinetechnique to optimize the impact of illumination on overall face recognition system. It is newlineimplemented on FPGA to ensure it is suitable for real time applications. We demonstrate newlinethat proposed algorithm outperforms over many existing algorithms and it is tested with newlinebroader set of databases. newline