- •1. Introduction
- •2. Proposed retinal identification algorithm
- •2.1. Vessel Extraction
- •2.1.1. Preprocessing
- •2.1.2. Vessel Segmentation
- •2.2. Regions of Interest Definition and Extraction
- •2.2.1.3. Angle of sr Corners
- •2.2.1. Feature Extraction
- •2.2.1.1. Region-based Shape Features
- •2.2.1.2. Boundary-based Shape Features
- •2.2.1.4. Centroid Distance
- •2.2.1.5. Differential Tangent Angle
- •2.2.1.6. Weighted Corner Angle Feature
- •2.3. Proposed Hierarchical Matching Structure
- •2.3.1. Candidate Selection
- •2.3.2. Corner Matching
- •2.3.3. Sr Matching
- •2.4. Decision Making Scenario
- •3. Experimental Results
- •3.1. Database Description
- •3.2.1 Utilized Feature Order in Candidate Selection Step
- •3.2.2 Number of the Most Similar Candidate sRs
- •3.2.3. Parameter Setting of sr Matching
- •3.2.4. Parameter Setting of Decision Making Scenario
- •3.3. Identification Result Analysis
- •3.4. Discussion and Comparison with Other Works
- •4. Conclusions
2.3. Proposed Hierarchical Matching Structure
The matching stage is the most challenging and time consuming stage
in all of related works, due to the fact that the size of feature
vectors can be varied because of the expressed issues and problems.
As previously remarked none of extracted features is completely
robust against rotation and translation issues but the manner we
engage these features in our hierarchical matching structure, makes
our algorithm robust against mentioned issues. Therefore, the
selection of matching scenario is a vital key to compensate these
problems and challenges. Hence, we employ a modified hierarchical
matching structure to match the query SR with the enrolled SR in the
database. Fig. 7 shows block diagram of proposed hierarchical
matching structure. This structure is composed of three modules such
as candidate selection, corner matching and SR matching. In each
module of this structure, we investigate the similarity of selected
SR of query image to the SRs of all enrolled images in the database
and some candidate SRs are removed from the search space in each
module of this structure. After applying candidate selection module,
the
candidate
SRs are selected so that in the corner matching module, the most
similar enrolled SR with query SR is determined. After that, the
final matching between selected enrolled SR and query SR is
performed. Finally, the final matched results from the SR matching
module are applied in the decision making scenario to identify or
reject the query image.
Fig. 7: Block diagram of proposed hierarchical matching structure.
2.3.1. Candidate Selection
As previously mentioned, the shape features are used to describe the SRs of retinal images. In the proposed hierarchical matching structure, we use these features to filter possible candidates in each stage of this structure. The sequence of employed features for filtering the candidates is relied on their abilities and efficiencies. In other words, the more complex and detailed features are applied in the higher steps to compare the SR query with the SR of candidates. In this module, we compare the query SR with the enrolled SR in database using only the region-based shape features.
As illustrated in Fig. 7, suppose
be a selected SR of the query image as the input of our matching
structure. We investigate the similarity of
in multiple steps with computing Euclidean distance between
and the candidate SRs in database. In each step, the most similar
candidates are selected to investigate with the query SR based on the
next feature in the next phase. Hence, the sequence of employed
features in the candidate selection module in Fig. 7 is considered
from more general and simpler features through more detailed and
complex features.
In each step of the candidate selection module, a number of samples of the most similar candidate SRs among database are selected to feed the next step in this module. Therefore, in each step, we eliminate a constant percentage of remaining most similar candidates from possible search space so that in final step M number of most similar SR candidates will remain for corner matching module.
