Research On The Algorithms Of Color Image Based Iris Recognition

With the fast development of science and technology, the demand of the people for security is increasing. The traditional security methods like key, password etc. are now facing more and more challenges, and they have the shortcomings such as not portable, prone to be forgotten and cracked, and so on. All these provide opportunity for the development of the intelligent identification methods based on biometrics, and they are receiving increasing attentions. Among them, iris recognition, because of its uniqueness, high reliability and non-intrusiveness, has been thought to be one of the most prospective biometric features, and it has now become a hot topic.Most of the previous iris recognition systems are based on the gray level images acquired from near Infrared cameras, with which the environments are strictly constrained, and the users are asked to cooperate with it, so the obtained images are very clear, as a result, the accuracy is very high. When applied to such sites like airport, custom and bank, these restrictions on the environments and users can be satisfied. But these constraints also limit the further application of this technique. In order to evaluate the performance of iris recognition based on color images from real conditions, this paper exploited UBIRIS color image iris dataset to study and develop the related algorithms, and analyzed their performance.This paper proposed a framework for color image based iris localization. The color image is first decomposed to several color channels, i.e., red, green and blue. After that, the inner and outer boundaries of the iris are localized in the red and blue channels respectively according to their characteristics. Finally, the results are combined together, and the area between the inner and the outer boundaries is the iris area.This paper presented two improved iris localization algorithms based on single color channel. The first algorithm calculates the threshold for the binarization of the image using gray level histogram, and then filters the binary image using morphological operators. After that, the intensity projection method is used to coarsely search for the center and radius of the inner boundary. Finally, Daugman's circular detection algorithm is employed in the binary image to finely localize the inner boundary. The use of binary image but not the gray level one has the advantage that it can reduce the influence of the reflections inside the pupil. And the coarse localization of the inner boundary can decrease the searching range, and thus speed up the system. On the basis of the detected inner boundary, when localizing the outer boundary, we improved the Hough transform algorithm with respect to the speed of the system. The system can be more than 3 times faster than before according to the points detected by the Canny edge detection algorithm. The second iris localization algorithm improved the algorithm of the localization of the inner boundary. Based on the character that, in most cases, the pupil is the darkest, smoothest circle in an iris image, it first finds the candidate circles using Canny and Hough transform, and then reduces the range gradually, and the last left candidate is thought to be the inner boundary.To minimize the influence of eyelid and eyelash, this paper proposed a simple algorithm to remove these areas based on a statistic method. After the outer boundary is detected, the system selects certain areas which are least possible to be eyelash or eyelids from the left and right parts of the circumference, respectively, and then count the range of the gray levels in these areas. The areas whose gray levels are not in that range are thought to be eyelids (whose intensity is below the lower bound) or eyelash (whose intensity is above the upper bound).The first work in this paper is single color channel based iris recognition. This paper analyzed the performances of iris recognition based on thirteen different color channels from five commonly used color spaces. The experiments show that, based on the feature extraction method using Log-Gabor filter, the Red, Y, I, Cr and r color channels provide perfect performances, while the channels like H, g and Q are not suitable. For the blue eyes, the accuracies of the green and blue channels are near to that of the red channel, and the accuracies of most of the channels from the blue eyes are above 95%.On the basis of the analysis of the single channels, this paper introduced matching score level data fusion methods to iris recognition to fuse the matching scores yielded by different color channels, using strategies like Min, Max, Sum, Product, User Weighting and Matcher Weighting. We studied the performances of the fusion strategies by fusing channels from the same color space, as well as those of fusing ones from different color spaces. It is found from the experiments that Max can not improve the performance in this case. The fusion strategies Sum, Product, User Weighting and Matcher Weighting are more robust than Min and Max. When applied to fuse independent information such as Y and Cr, these fusion strategies can increase the accuracy of the system. But if they are applied to correlated information such as R, G and B, only when the accuracy of these channels themselves are high enough and near to each other, can they provide higher performance than the single channel.This paper proposed a decision level data fusing algorithm which combines weighted k nearest neighbors (WKNN) with weighted majority voting (WMV) methods for iris recognition. First, find k nearest neighbors of each channel and set the weights for them according to the distance measurements between it and the corresponding channel of the templates, in which three distance calculation methods are employed. Then, vote for the class the image belongs to by weighted sum of the weights of all the channels, in which the weight of each channel is determined by its performance measurements such as accuracy, EER, mean and covariance. It also presented an algorithm for iris recognition, which is based on the combination of weighted fuzzy k nearest neighbors and weighted majority voting. Experiments show that these methods can reliably improve the accuracy, overcome the disadvantage of the traditional KNN algorithms, effectively reduce the influence of noise, and enhance the robustness of the iris recognition system. Additionally, this paper analyzed the rule of the accuracy changing with K when it varies from 1 to 10.Finally, this paper studied the performance of iris recognition based on multiple templates, and presented semi-supervised learning and incremental learning algorithms for iris recognition. In the mode of incremental learning, the managers update the templates for each class periodically or aperiodically, and the system can learn from the newly added templates. While in the mode of semi-supervised learning, the system can automatically update the templates during authentication of identification, because it has the ability to study from the input images. The experiments show the effetiveness of these methods.The works above can provide valuable reference to the future research on color image based iris recognition.Though the proposed methods can ameliorate the performance of the iris recognition system, there are still many shortcomings in this paper. This paper didn't employ any image quality evaluation measurement to discard useless images or provide reference to the future processing. The feature extraction methods in this paper are all based on single color channel, which result in the fact that color information - the relationship among the different channels is not effectively employed. So we should also consider the feature extraction and matching algorithms based on color information. In the meanwhile, though some of the channels are proved not to be suited to Log-Gabor filter based iris recognition, they could provide useful helps in some other respects. How to make full use of these color channels is also a problem worth studying. All the problems above, together with iris live detection, are the emphases of the future work.