This paper examines the possibility of predicting ethnicity based on iris texture. This is possible if there are similarities of the iris texture of a certain ethnicity, and these similarities differ from ethnicity to ethnicity. This sort of “soft biometric” prediction could be used, for example, to narrow the search of an enrollment database for a match to probe sample. Using an iris image dataset representing 120 persons and 10-fold person-disjoint cross validation, we obtain 91% correct Asian / Caucasian ethnicity classification.
Iris texture has been shown to be useful for biometric
identification and verification
The CASIA biometrics research group has performed
research on iris texture elements, including studies
In a study of how human observers categorize images,
Stark, Bowyer, and Siena (2010) found that humans
perceive general differences in iris texture that can be used
to classify iris textures into categories of similar texture
pattern. Observers grouped a set of 100 iris images into
categories of similar texture. The 100 images represented
100 different persons, and the 100 persons were balanced
on gender and on Asian / Caucasian ethnicity. The
observers did not know the gender or ethnicity of the
persons in the iris images. However, the grouping of
images into categories of similar iris texture resulted in
categories that were, on average, split 80% / 20% on
ethnicity. The same categories were on average divided
much more closely to 50% / 50% on gender. Thus, one
result of Stark’s work (2010) is that human observers
perceive consistent ethnicity-related differences in iris
texture. In this paper, we want to train a classifier to
explicitly perform the sort of ethnicity classification that
was found as a side effect of the texture similarity grouping
done by humans in
We want to see how accurately we can identify ethnicity
based on iris texture. For this study we will use two
ethnicity classes, Caucasian and Asian. This study used
1200 iris images selected from the University of Notre
Dame’s iris image database. (This is a newer database than
was released to the iris biometrics research community for
the government’s Iris Challenge Evaluation (ICE) program
For this iris texture prediction study, we want to base our findings solely on iris texture. Therefore we exclude periocular clues that might be used as an indicator of ethnicity. We segment the images to obtain the region of interest, and mask out the eyelid-occluded portions of the iris. We use Notre Dame’s IrisBee software to perform the segmentations (Phillips et al. 2005). The output from IrisBee that we use for texture examination is a 240x40 pixel normalized iris image along with the corresponding bitmask of eyelid and eyelash occlusion locations. The image segmentation and masking are exactly those that would be used by IrisBEE in processing the images for biometric recognition of a person’s identity. However, the normalized images are not processed by the log-Gabor filters that are used by IrisBEE to create the “iris code” for biometric recognition. We create a different texture feature vector for ethnicity prediction.
After an image is segmented and normalized, we compute
texture features that can be used in training a classifier to
categorize images according to ethnicity. To do this we
apply different filters to the image at every non-masked
pixel location, and use the results of the filter to build a
feature vector. Six of the filters we have used are “spot
detectors” and “line detectors” of various sizes, as depicted
in Tables I to VI. For a given point in the image, if
applying a given filter would result in using any pixel that
is masked, then that filter application is skipped for that
point. The rest of the filters, depicted in Tables VI-VIII,
were created using Laws’ Texture Measures
A feature vector that describes the texture is computed for each iris image. We divided the normalized image array into a number of smaller sections in order to compute statistics for sub-regions of the normalized image. This is so that classification could be based on, for example, relative differences between the band of the iris nearer the pupil versus the band of the iris furthest from the pupil. These regions were ten four-pixel horizontal bands and four 60-pixel vertical bands of neighboring pixels in the normalized iris image. The ten horizontal bands correspond to concentric circular bands of the iris, running from the pupil out to the sclera (white) of the eye. The four vertical bands correspond roughly to the top, right, bottom and left parts of the iris. Since the filters are looking for different phenomena in the image, we find statistics for the filter response of each image. Each image contains 630 features, with 5 statistics calculated for each of the 9 filters on all of the 14 regions. The five statistics are: (1) average value of filter response, (2) standard deviation of filter response, (3) 90th percentile value of filter response, (4) 10th percentile value of filter response, and (5) range between 90th and 10th percentile value. The motivation for using the average value is to represent the strength of a given spot size or line width in the texture. The motivation for using the standard deviation is to represent the degree of variation in the response. The motivation for using the percentiles and range is to have an alternate representation of the variation that is not affected by small amounts of image segmentation error. Results We tried a variety of different classification algorithms included in the WEKA package (Weka). This included using meta-algorithms like Bagging with other classifiers. By changing parameters, we achieved performance gains on some of the algorithms. However, we found our best results using the SMO algorithm with the default parameters in WEKA for classification. The SMO algorithm implements “Sequential Minimal Optimization”, John Platt’s algorithm for building a support vector machine classifier (Weka). The input to the SMO algorithm is the feature vectors of all 1200 iris images that we have computed. To assess the results of our classifier we use cross-fold validation with ten folds using stratification based on ethnicity. These folds are also subject-disjoint to ensure the persons whose images are in the test data have not been seen by the classification algorithm in the training data.
The SMO classifier results in higher accuracy compared
to a broad range of other classifiers, including decision tree
based algorithms and bagging. Using Bagging on the top
two classifiers, SMO and Random Forest, did not improve
performance. Running the experiment with the SMO
classifier and the feature vector as described above gives
us an accuracy of 90.58%. This is good accuracy,
representing an improvement on the 86% reported in
We computed the classification accuracy for each feature separately to see the impact of individual features. Table X shows that some of the single features have almost have the performance of all of the features together. However none of them do as well as the combination of all of the features. Some filters may be redundant; a combination of a few might reproduce the performance of all nine filters.
To ensure that the size of our training dataset was not limiting our accuracy levels, we ran the classifier with different numbers of folds. Table XI shows the results we achieved using 5, 10, and 20 fold cross validation. The accuracy levels are all within one percent, indicating that our performance should not be limited by our dataset size.
To achieve even greater accuracy, we intend to implement additional and more sophisticated features, and to look at the effects of the size of the training set. We envision that the number of different persons represented in the training data is likely to be more important than the number of images in the training set; that is, doubling the training set by using twice as many images per person is likely not as powerful as doubling the number of persons.
For this experiment, we only looked at very broad ethnicity classifications. More work could be done to examine finer categories, such as Indian and Southeast Asian. The performance of a classifier such as this has not been tested on subjects of multiple ethnic backgrounds either.
This work is supported by the Technical Support Working Group under US Army contract W91CRB-08-C-0093, and by the Central Intelligence Agency. The opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of our sponsors.