The Research Of Critical Technology For Face Recognition Under Complex Lighting Conditions

Face Recognition has the advantages of initiative, user-friendliness andnon-invasion. Due to its enormous social value and application prospect in the field ofnational and public security, information and financial security, and human-computerinteraction, face recognition has been a hot research field of pattern recognition andmachine vision. Although face recognition has numerous applications in many areas ofpeople’s lives, the research of face recognition still faces enormous challenges. A lot offactors can affect the performance of face recognition, among them, light is the primarychallenge. Since light pre-processing and feature extraction are the main bottleneck ofsolving illumination problem, this thesis focuses on the research of light pre-processingand feature extraction algorithms, which based on the research and analysis of existingrelated algorithms.In the aspect of light pre-processing, existing related algorithms, especially thealgorithms on basis of Retinex theory, are researched and analyzed, and then a new lightpre-processing algorithm based on the optimal conduction function and Retinex theorywith adaptive smoothing ability is proposed. By describing inconsistencies of the faceimage pixel points, illumination is estimated via adaptive smoothing filter, and then thenatural facial features are restored. Experiments show that the algorithm proposed canachieve better result than classic Retinex algorithm, especially can effectively identifyface images under strong-side lighting conditions.In the aspect of feature extraction, take account of that texture features arerelatively insensitive to illumination changes, LBP(Local Binary Patterns) and itsextended algorithms are discussed, and then a new texture feature extraction algorithm,named Complete Local Ternary Patterns, is proposed. This method extends thetraditional LBP from the direction of the gray scale differences as well as differences inmagnitude, etc. It can extract texture features that include higher discrimination and canalso overcome the traditional binary pattern’s coding inconsistent problem onnon-uniform illumination. Experiments show that the algorithm proposed can extract more texture information than the traditional binary pattern, and can achieve betterrobustness for slight non-uniform illumination at the same time.