In the decision making scenario, to identify an individual in the database, two matched SRs of the query image and the same individual should be satisfied the introduced measure in Eq. (15). In this equation, the threshold determines the level of identification and rejection of the individuals. To select the optimal value this threshold, we examine our decision making scenario for different values of in an experiment similar to experiment A with different rotation and translation values while other parameters have been set as explained previously. Fig. 10 shows the identification results in terms of FAR and FRR for various values of . In Fig. (10-a) the lowest error rate occurs in with FAR and FRR of zero. Note that in this experiment, the FAR and FRR devoted to the individual identification not the SR matching. Again to reduce the error of FRR, we consider where the FRR is equal to zero with a high confidence limit. In Fig. (10-b), FAR for all threshold values are equal to zero. This fact is due to our applied strategy during the hierarchical matching process and specially in our decision making scenario. It is technically impossible to match mistakenly two SRs belong to a query image with two SRs belong to the same individual in dataset while the distance between centroids of these SRs is the same both in the query image and the enrolled image. Although in our decision making scenario, false reject still could happened in cases that the SRs are not extracted, but these cases occur mostly in the retinal images of diabetic patients with severe damages in the retina. However, in all stages of parameter tuning of the hierarchical matching structure, we tried to select the optimal values based on minimizing the false reject so that both FAR and FRR be zero.
(a)
(b)
(c)
Fig. 10: Identification results of our proposed decision making scenario for different values of based on ROC curves. (a) Error rate, (b) FAR/FRR, (c) Genuine accuracy versus FAR.
After setting two thresholds and with the optimal values, to evaluate the performance of our proposed identification algorithm precisely, we tested the proposed algorithm in an experiment similar to experiment A with different rotation and translation values. Fig. (11-a) shows the number of individuals that have been identified with different numbers of the investigated SRs in the decision making scenario. Obviously, the most of images (i.e. 140 images) are recognized by investigating their first two SRs. For the query images which are identified with three investigated SRs, surely one of the first two investigated SRs has been rejected and in the case of four investigated SRs, two SRs of the query have been rejected. This fact is not necessarily due to mistake in the matching structure while new SRs may be created or disappeared because of rotation and translation processes or different illumination conditions. For the query images which are identified after investigating ten SRs, at least one of the first nine SRs have been matched mistakenly with one of enrolled SRs in database.
In the images that all ten extracted SRs are investigated, the matching time reaches to the maximum value. Fig. (11-b) shows average of the processing time for each individual concerning this experiment and accuracy rate of identification by investigating different numbers of the extracted SRs. In this figure, with increasing from 1 to 5 the accuracy rate of matching is improved and as a result the time of matching procedure also is increased. But, with increasing from 5 to higher values the accuracy rate of matching does not change while the matching time is increased. In other words, when we investigate just the first five largest SRs for each image, the operating time is pretty low while the identification performance is still high. In order to reduce investigating time in matching process, if the first five largest SRs of the query image are rejected in the matching step then the query image is rejected. Therefor instead of investigating nine SRs and then rejecting the query we reject the query while its first five SRs are rejected consecutively. The result of this change is evident in experimental results in Fig. 12.
(a)
(b)
Fig. 11: Results of evaluation the proposed hierarchical matching structure. (a) The number of individuals that have been identified with different numbers of the investigated SRs in the decision making scenario, (b) Average of the processing time for each individual and accuracy rate of identification by investigating different numbers of the extracted SRs.