The Research Of Pupil Precise Localization Method

Pupil location is an important research issue in the calculator sense of vision field involving various disciplines, such as the physiology, psychology, artificial intelligence, pattern recognition, computer vision, image processing and analysis, etc. The pupil location technique first develops along with the development of face analysis technique such as face examination, facial expression analysis. In the 60's and 70's of the last century, face recognition has been put forward, thus need of pupil examining is urgent. This urgent need of pupil location technique has been well satisfied with the further development of face analysis and iris recognition.The research aim of this topic is to study the computer vision, the image processing; the pattern recognition and other related knowledge as well as the combination of them on the basis of the existing research results in order to search for a new method of pupil location. The focus of the study is put forward an improved method of pupil location, which is based on symmetry wavelet, grads transform and symmetry transform. Specific details are as follows:1,Introducing some basic knowledge of image processing and existing pupil location methods, for example: Hough transform, deformable template, edge characteristic analysis, Geometrical face model, symmetry transform, etc.2,Researching the theory and its realization of wavelet transform. The face image of high frequency information is more concentrated in eyes, mouth and pupils which are strongly symmetrical. High pass filter to the image by symmetry wavelet has been carried out through comparing experiments. Experimental results show that the symmetry wavelet high pass filter can effectively highlight the high frequency information and provide a good foundation condition for the follow-up work.3,Putting forward a new method to choose the effective region of eyes. The biggest problem of pupil location through symmetry transform method is large numbers of calculation, therefore it is particularly important to put up regional choice before symmetry transform in order to reduce the number of pixels for calculation. The paper put up mean filter of 5*5 template to the wavelet filter image. Through several experiments, receive threshold image by five times of the average value of the pixel vie the principles of mathematical morphology. Choose the dilatable factor which elements structure radius is 5 pixels carrying out dilation operation; finally, choose the effectual region by implementing and operation between the dilatable image and original image. Experiments show that this method can choose the effectual region of the image to process so as to greatly reduce the latter part of the symmetry transformation calculation as a result.4,Improving Generalized Symmetry Transform method. Analysis shows that it is an opposite direction comparing the human eye pupil region with reflective glasses region. It only gets the maximum in the pupil region because of improvement operator to better resolve the problem of the glasses'reflection. There are often a few high value pixels in the pupil's low gray area (such as reflection) in the grey image. Therefore pupil region has a poor symmetry of round, but the edge of the pupil has a strong symmetry of circle. The operator divides eyes region into circular symmetric region and oval symmetric region. Take into account the eyeball which may be covered by eyelid, the restrictive condition yi ? y j< 2(σ0+ 1) is introduced. The experiments show that by using this method, the large numbers of strong symmetric point by calculation mainly concentrate in the eye region while just a few concentrate in the eyebrow region and fewer appear between the nose and mouth.5,Putting forward one Self-evaluation based on the strong symmetry of the eyeball region, the similarity of two eyes and the consistency of gradient direction of link between two pupils. The paper defines the symmetry sym ( p )' and angle consistency S; respectively intercepts the left eye and the right eye image and overturn it at the same time; defines the first comparability cor1 , the second comparability cor2 ; set the threshold value, compares it with cor1 , cor2 and S, judges the location accuracy.6,Dividing the points which have the highest symmetric value into a, b, c three groups to mark the location point according to the result of grouping. (1) All candidates belong to a. This situation only accurate locates one pupil; in a determinate region, choose the point which is the most similar to the point P1 as the result of anther.(2) Candidates belong to a, b and c respectively In such circumstances, group c often appears in the mouth or between nose and mouth. General can use relative position relation to get the candidate locate on the pupil.(3) Candidates belong to a, b respectively.It can be divided to two situations:①P1 1 or P2 1 exist: Possible is eyes closed or bushy eyebrows. In this circumstances, P1 1 or P2 1 is considered the accurate pupil, then process it as the first situation (all candidates belong to a).②P1 1 and P2 1 not exist: According to the self-evaluation to determine whether accurate location, if accurate, then skip the second location, output location result directly; or else, calculate directional symmetry, if sym(P1 )'≥1.4sym(P2 )' , P2 is accurate location, otherwise, P1 is accurate, then process it as the first situation (all candidates belong to a).