The Study On Face Real-time Detection Based On Adaboost Algorithms

In recent years,with computer technology in the field of human-computer interaction rapid development of intelligent information processing as a key technology,face detection has become the pattern recognition,computer vision and human-computer interaction field of a universal importance and very active research topic.Including video telephony,video conferencing,intelligent computers,multimedia intelligent entertainment,security authentication and monitoring system has many aspects,such as a wide range of applications. At the same time,face detection technology breakthrough will be to face recognition, facial expression gesture recognition, video surveillance, authentication,multimedia databases,and many other research in the field of a great role in promoting,but also in the methodological point of view the promotion of computer vision, pattern recognition,such as computer science or even a whole branch of computer science development.Thus,the study of face detection technology has great academic significance and application value.Literature interiorly and abroad,from the detection face of the rules used one can be divided into face detection method based on a priori knowledge and a posteriori-based learning and training methods.Learning-based face detection methods,the three most important factors:face modeling,the choice of learning algorithm and classifier structure.Paul Viola and others to build real-time face detection system,although a good performance,fast speed,can achieve 15 frames/sec.However,there are still some issues to be resolved,and Paul Viola did not give a specific method.As an exploration and attempt,this paper,a number of methods of Paul Viola to expand and improve, to achieve a positive real-time face detection.Characteristics of the object in need of classification of information has been encoded.The face image,the more favorable characteristics than the pixels on the face of that,but the information from the original pixels in the image difficult to obtain.Characteristics of the final results of learning algorithms have a great impact, mainly in the accuracy and speed of both the impact of speed to clear. In terms of characteristics,characteristic values of the computational complexity is also a very important aspect.This is because almost all of the object detection algorithm must scan a large number of windows, each window must be calculated for the corresponding eigenvalue.Harr-like feature of face detection technology in the development of an objective found a way to model objects.This feature of simple form that is conducive to the rapid calculation,at the same time in various forms.In this article,Paul Viola methods we used in the Harr-like features of the expansion,Rainer Lienhart,such as the use of the proposed expansion of the rectangular characteristics.Compared to the rectangular characteristics of the Treasury Paul Viola used in the case of the rectangular characteristics, mainly an increase of 45°rotation of the rectangular characteristics.Reuse integral images(that for a rectangular area on the rapid integration of a discrete function of the data structure)the concept of each image to be detected,as well as all the training samples can be images of the points by calculating the corresponding map,find its rectangular eigenvalue,and only once.Thus expanding the scope of training to improve the detection rate and lower rate of false.that,we know Adaboost learning algorithm of the basic idea is:given a weak learning algorithm and a training set: {( x1 , y1 ),( x2 , y 2), ???, ( x N , yN)},which xi is the input vector of training samples, yi is marked category.Face detection task is to determine whether the image is a face image,it can be seen as the distinction between two types of problem, you can take yi∈{0,1},where 0 and 1,respectively, and are cases of counter-examples.In the initialization of all training samples are given to the same weight,and then use the weak learning algorithm on the training sample set T round of training.At the end of each round of training,the training samples failed to give greater weight to allow the subsequent learning algorithm in the main contrast is more difficult to learn the training samples for learning.This can be a weak classifier sequence: {h 1 , h2 , ???, ht},where the effect of a better classification of the weak classifier greater weight.The final classifier H ( x ) uses a re-vote the right way.Based on the above learning algorithm Adaboost-depth analysis and research, Masayuki Nakamura et al in the study,based on the Adaboost learning algorithm of the original value of the right to update the rules to do a number of improvements:the weight of each sample is treated as a judge of the It is difficult to sample a target sample,and,assuming that the power had been given the value of the samples are difficult to sample.If a sample was correctly classified,then its weight will be in accordance with the original Adaboost learning algorithm to update the rules,namely:to reduce the samples correctly classified weight.If a sample was incorrectly classified,then the current to the sample weight,compared to HighWeightt :If the sample of the current weight is less than HighWeightt ,weight of the samples will be increased.If the sample of the current weight is greater than HighWeightt ,weight of the samples will be reduced.In this way,even if it is difficult in each round of samples were wrongly categorized,their weight will not be increased too much,thus to some extent,the classifier to avoid learning the phenomenon occurred.In order to ensure the detection rate based on the lower rate of false, Paul Viola Face detection in the use of multi-layer classifier cascade (Cascade Classifiers).This multi-storey structure is in fact a degradation of the decision tree.For an image to be detected, it is the first to be sent to the first layer classifier (at every level of the strong classifier by AdaBoost method of training to be),if the image was not a non-human face,the end of the testing process,or else be classifier to the second tier,and so on.Classifier is only through all the images will be considered to be face image.Therefore,this multi-layered structure of the advantages is that,through the integration of multiple classifier role can significantly reduce the rate of false and at the same time due to the classification of multi-layer structure,and makes the majority of non-human faces in relatively low-rise the top of the classifier has been ruled out,taking into account the detection of human faces to be detected when the majority of non-human face images are images in this,kind of structure to ensure a higher rate of detection.Through Face Detection of experimental results point of view:The AdaBoost algorithm into the background image in a simple,different face in different lighting conditions,different facial expressions and facial images of different sizes,the performance of a high The detection rate and detection rate,with better robustness.Perspective in a multi-face detection,the study compared the current and more effective methods of face detection perspective,and positive in front of face detection algorithm based on the classifier using the pyramid thinking, for out-plane [-90°,90°] of face training rotation n rotation direction different point of view of the classifier.Through the pyramid's point of view the classification decision,and finally to build a multi-angle face detection framework.In this paper,from a practical point of view in order to improve the accuracy and real-time for the purpose of expanding the application of the rectangular characteristics of Adaboost learning algorithm is also made improvements to achieve a positive real-time face detection; and learn from Stan Z.Li pyramid Category ideological device for multi-angle face detection algorithm is studied.