Progressive Iris Centers Localization Method Based On Double Cascaded Regression

Accurate iris center localization is an important pre-step for applications such as gaze track-ing,human-computer interaction,attention detection,and iris authentication,thus has a wide range of applications.The eye images captured by the camera in the daily scene are often affected by factors such as filming angle,illumination change,lens reflection,etc.,and thus leading to image noise or iris occlusion.Therefore,accurate iris center localization for daily applications still has great challenges.For the requirements of a wide range of applications,this paper focuses on the key techniques of the iris centering method based on image processing,such as the extraction of the eye region,the localization of the iris center and the refinement of the iris center.The main research contents of this paper are as follows:1.Eye region extraction is the basis of the iris center localization algorithm,whose accuracy greatly affects the performance of the iris center localization method.Based on the research and analysis of predecessors related methods,this paper establishes an eye region extraction method which based on the detection of canthi and human eye anatomy proportional relationship.Taking the canthi detected by the cascading regression forest as the reference points,the deviations be-tween the eye region boundaries and the canthi were adaptively estimated according to the human eye anatomy proportional relationship,thereby extracting an accurate binocular region.The per-formance of eye region extraction was verified on the public database BioID,and 1514 samples were able to extract the eye region containing complete binocular among 1515 samples which detected the face.2.The paper studies and summarizes the existing methods and related theories of iris cen-ter localization.Based on the analysis of the predecessor methods,the multi-stage scheme was employed to combine the advantages of various methods and progressively refine the iris center.The cascading regression forest model,which is robust to head posture changes,was chosen as the basis of the iris center localization algorithm and the shape used for cascading regression was redesigned.The trained cascading regression forest was employed on the eye region image to detect the rough iris center and the landmarks on the eye contour.The open or closed state was determined by the contour of the eye,which determines whether further steps performed to refine iris centers.For the open eyes,the estimated iris point set was extracted by using the eye contour and the coarse iris centers.It can eliminate the interference and reduce the amount of calculation,thus ensures that the subsequent steps arc performed on a reliable basis;According to the grayscale characteristic of the eye,the inversed intensity weighted average centroid of the iris point set was calculated to produce the new iris center estimate;According to the strong intensity difference between the iris and the sclera of the eye,the snakuscule model was used to evaluate the quality of the iris center estimation.The radius iteration is removed from the original snakuscule model,and the weighted approach was used to improve the original snakuscule model.The weighted snakuscule model was used to refine those unqualified iris centers.3.The relevant parameters in our proposed method were verified and optimized by ex-periments,and the performance of iris center localization was verified on the public database BioID,GI4E,Talking Face Video.The performance of the algorithm was analyzed from the as-pects of real-time and accuracy.On the challenging BioID database,the algorithm performs at 33 frames per second including face detection,which meets the requirements of real-time appli-cations.Moreover,the localization accuracy of our proposed method on the BioID database is 74.69%and 96.58%under the standard of normalized error e<0.025 and e<0.05,respectively,which exceeds the accuracy of the previous method.