2.4. Decision Making Scenario

To identify or reject the individuals concerning the matched SRs in the proposed hierarchical matching structure, we propose a decision making scenario based on the matching results of SRs. This decision making scenario is defined so that the identification or rejection of an individual with the least number of SR matching process is carried out. For this purpose, according to matched SRs of the query in each irritation, decision making scenario determines that the SR matching process should be continued or stopped.

Assume and be the first two matched SRs (SR a and SR b) that belong to the same individual in the database and , be the centroids of these two SRs. Moreover, and be their corresponding matched SRs in the query image with the centroids of and . Euclidean distance between the centroids of each pair of these matched SRs in the query and enrolled image in the database is computed as,

(14)

Since the centroid of SR is the most robust point of a SR against noise and movement factors in whole of a retinal image (because the centroid is obtained by averaging all points of a region and not only border points), we use the distance between the centroids of two SRs in the query and enrolled images in the database to decide in the decision making scenario as in the following,

(15)

where is a predefined threshold. In Eq. (15), if individual is identified, the SR matching process will be stopped. Otherwise, the SR matching process continue until our decision making scenario identify an individual image in the database or all greatest extracted SR from the query image are investigated and finally the query image is rejected if none of the matched SR pairs does not satisfy the condition in our decision making scenario(Eq. (15)). Fig. 8 shows a query image that has been identified by an enrolled image in the database. In the images of Fig. (8-a) and Fig. (8-b), two pairs of the matched SRs in the enrolled image in the database and query image are shown with their centroids. Notice that the distance between two centroids described in Eq. (14) is completely different from centroid distance explained in section (2.2.1.4) and used in corner matching (section (2.3.2))

(a) (b)

Fig. 8: Decision making scenario. (a) Two matched SRs in an enrolled image of the database with their centroids, (b) Two matched SRs in the query image with their centroids.

3. Experimental Results

3.1. Database Description

To evaluate our work, proposed algorithm is applied to two databases composed of 40 color retinal images (565×584 pixels) from DRIVE database [35] and 20 chosen color retinal images (720× 576 pixels) from STARE database [36] which are both publicly available. We rotated and translated each subject randomly in different orientations and directions five times to obtain 200 images from DRIVE database and ten times to obtain 200 retinal images from STARE database. To implement our proposed algorithm, we used MATLAB software (version 7.14, R2012a) on a PC with CPU, Intel Core i3 and 2MB RAM, in six experiments as in the following.

3.1.1. Experiment A

The first 30 retinal images of DRIVE database were taken as enrolled images and the last 10 images with five images for each subject were considered as query images.

3.1.2. Experiment B

The last 30 retinal images of DRIVE database were enrolled and the first 10 images plus five images for each subject were considered as query images.

3.1.3. Experiment C

The first 15 images plus the last 15 images of DRIVE database were considered as enrolled images and 10 rested images of DRIVE database with 5 images for each subject were taken as query images.

3.1.4. Experiment D

All 40 images of DRIVE database were considered as enrolled images and 200 translated and rotated images were entered as query images. Obviously, in this experiment none of query images is rejected.

3.1.5. Experiment E

The first 15 selected images of STARE database were taken as enrolled images and 200 rotated and translated images were considered as query images.

3.1.6. Experiment F

The last 15 images of STARE database were considered as enrolled images and 200 rotated and translated images were considered as query images.

3.2. Parameter Setting

To select and set the optimal parameters in the hierarchical matching structure, we examine some experiments with an experiment similar to experiment A with different rotation and translation values. In these experiments, the issues such as the selection of utilized feature order in the candidate selection stage, the number of the most similar candidate SRs among database ( ) that are determined by the candidate selection stage, the optimal selection of thresholds in the SR matching stage and decision making scenario should be determined. In the following, the results of these experiments will be described.