The Research On Algorithms For Face Detection And Recognition Under Complex Environment For Robot

Face detection and recognition technology has been applied in many fields,and compare face with other biometrics,the advantage of face detection is non-contact.The robots,especially with the visual system,need to verify the identity of their‘master'in front of it,and now with more is face recognition.But also meet various interference of environment for the character of non-contact,in addition,also meet interference of posture,expression changes,which will also affect the final recognition result.The dissertation mainly studies the algorithm of face detection and recognition for robots,which consists three parts including the algorithm of illumination compensation,the design of face detection system,the design of face recognition algorithm.At the side of illumination compensation,the thesis improves the algorithm of Retinex,and only enhances the luminance component of color image by MSR,and uses equal proportion transformation to restore the color.Experiments show that in image enhancement aspect,our algorithm has obvious advantage compared with other algorithms referenced in this paper,and maintains color information while making illumination compensation.At the side of face detection,the thesis designs two-stage detecting system with high detection rate and low false detection rate.The system uses cascade of classifiers based on Adaboost as the first level,and uses skin segmentation to check“human-like faces”.The novel system can eliminate most of the false detection areas,and has lower false detection rate compared with single algorithm.At the side of face recognition,the thesis first improves the algorithm of Local Ternary Derivative Pattern(LTDP~2),and proposes LATDP~2 with a novel adaptive threshold.Improve the universality of the original algorithm greatly.Then operate on the Gabor magnitude features with LATDP~2 to get higher identification of face description.Experiments show that the new algorithm compared with LDP~2 and LBP has better robustness under illumination and expression variation and the recognition rates also have a great improvement compared with other algorithms.