| Face recognition is a typical biometric identification technology, which is related toimage processing, pattern recognition, artificial intelligence and other disciplines, withhigh academic research values and rich business scenarios. Face detection is an importantpart of it. What’s more, eyes are the most important facial features. The location algorithmalso has a very high research value.This thesis describes several common image pre-processing methods, giving adetailed description of their calculation processes and displaying specific pretreatmenteffect diagrams, which has eliminated image noises and reduced the impact ofillumination changes on the identification work to a certain extent. Second, the thesisdiscusses the basic principles of classical Adaboost knowledge and its calculation processin face detection applications. After getting the degree of difficulty of identifying thesamples, we propose the concept of common samples, according to which we adoptdifferent weight update strategies to get more efficient classifiers, improving the detectionrate and reducing the false detection rate. The optimized Adaboost cascade structure in thisarticle also improves the detection results. The introduction of auxiliary functions hasimproved the human face detection accuracy in the premise of less detection time increase.In the human eye coarse positioning phase, we limit the search area at top2/3parts of theface, thereby reducing the amount of computation, eliminating part of the interference, andensuring that the eye area is not excluded. Finally, eye precise positioning is completedbased on the isophote curvature and gradient features. According curvedness and gradientsigns of points on the isophote, we select part of the displacement vectors to vote for theisophote center, followed by result correction using the mean shift recognition algorithm.Experimental results show that the classic Adaboost-based improved algorithm andisophote-based human eye location method have effectively improved the measurementaccuracy, achieving very good recognition results in the cases of profile, linearillumination changes and partial occlusion. The precise positioning of eyes is based onlow-dimensional features, with small amount of calculation to meet real-time location demands. |
