研究目的
To introduce a method for parameterizing CNN-based iris segmentation to enable its use in traditional biometric recognition systems, bridging the gap between CNN segmentation and the rubbersheet transform, and to evaluate its performance compared to traditional methods.
研究成果
The proposed parameterization method works well across tested databases with no failures. CNN-based segmentation outperforms traditional methods on lower-quality and difficult databases but is surpassed on high-quality NIR images with frontal acquisition. Using CNN segmentation as a noise mask improves traditional methods in some cases but not on very high-quality or very difficult recordings.
研究不足
The method may not perform as well on very high-quality iris images where traditional segmentation methods excel. It relies on circular parameterization, which might not handle elliptical shapes well in off-angle recordings. The CNN-based segmentation loses texture information used in traditional methods.
1：Experimental Design and Method Selection:
The study uses CNN-based segmentation (specifically RefineNet and iFCEDN) for iris segmentation, followed by a parameterization method involving preprocessing with median blurring, circular Hough transform for candidate generation, and selection based on value functions. The biometric toolchain includes feature extraction using wavelet-based (qsw) and log-Gabor (lg) features, and matching with USIT toolkit.
2：Sample Selection and Data Sources:
Databases used include IIT Delhi Iris Database (iitd), CASIA Iris Image Database version 4.0 Interval subset (casia4i), ND-0405 Iris Image dataset subset (ndi), CASIA Iris Subject Ageing Version 1.0 Database subset (casiaA), and PROTECT Multimodal DATABASE iris images (protI). Ground-truth segmentations are available for all databases.
3：0 Interval subset (casia4i), ND-0405 Iris Image dataset subset (ndi), CASIA Iris Subject Ageing Version 0 Database subset (casiaA), and PROTECT Multimodal DATABASE iris images (protI). Ground-truth segmentations are available for all databases.
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: No specific hardware mentioned; software includes USIT toolkit, RefineNet CNN, and datasets as specified.
4：Experimental Procedures and Operational Workflow:
Five-fold cross-validation is used for CNN training and testing. Segmentation performance is evaluated using type-1 and type-2 errors and F-measure. Biometric recognition performance is assessed using equal error rate (EER) and false non-match rate at false match rate of 0.01%.
5：01%.
Data Analysis Methods:
5. Data Analysis Methods: Statistical analysis includes McNemar test for significance; performance metrics include EER, FNMR@FMR=0.01%, segmentation errors, and masking errors.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
