Introduction
Deep learning models, particularly Convolutional Neural Networks (CNNs), have revolutionized biometric recognition. However, their performance is heavily dependent on the availability of large, diverse datasets. In iris recognition, acquiring such datasets is often constrained by privacy regulations and the technical difficulty of capturing high-quality images. This leads to the "small sample" problem, where models overfit to limited training data. While standard augmentation techniques like flipping or rotation are common, they can distort the intricate texture patterns unique to the iris. This paper proposes a novel intensity-based augmentation method using a modified Sparrow Search Algorithm (CPSSA) to optimize Contrast Limited Adaptive Histogram Equalization (CLAHE).
Methodology: The CPSSA-CLAHE Framework
The proposed method shifts the focus from geometric transformations to intensity-based variations. The core idea is to use CLAHE to enhance image contrast, but instead of using a static clipping limit, the algorithm searches for an optimal set of limits that generate diverse yet identity-preserving images.
Chaotic Pareto Sparrow Search Algorithm (CPSSA)
To navigate the search space of clipping limits efficiently, the authors developed the CPSSA. This metaheuristic improves upon the standard Sparrow Search Algorithm by incorporating fractional-order chaos for better initial population distribution and Pareto distributions to prevent the algorithm from converging on local optima.
As shown in the flowchart, the system extracts Log-Gabor features from generated images and calculates their cosine similarity to the original. This ensures that the augmented data maintains the biological characteristics necessary for accurate identification.
Experimental Setup and Datasets
The study utilized the CASIA-Iris-Thousand dataset, selecting 2,000 images from 200 individuals. To simulate a small-sample scenario, only five images per person were used for training.
Before augmentation, images underwent a preprocessing phase where the iris was localized and cropped (as seen in the figure above). These "separated images" served as the input for the CPSSA-CLAHE pipeline. The augmented datasets were then used to train four prominent CNN architectures: AlexNet, VGG-16, ResNet-50, and Inception-V3.
Key Results and Discussion
The CPSSA-based augmentation consistently outperformed both raw datasets and traditional geometric augmentation methods across all models.
- Accuracy Improvements: The Inception-V3 model achieved a peak accuracy of 95.50% when using the proposed method, compared to 89.83% with the original data.
- Error Rate Reduction: Biometric reliability, measured by the Equal Error Rate (EER), saw significant improvements. For ResNet-50, the EER dropped from 1.853% to 1.012%.
- Comparison with Traditional Methods: Interestingly, traditional geometric transformations like flipping or zooming were found to be counterproductive in some instances, such as with AlexNet, where accuracy actually decreased. This highlights the importance of preserving texture integrity in biometric data.
Conclusion
By treating data augmentation as an optimization problem, the researchers successfully addressed the small-sample challenge in iris recognition. The integration of swarm intelligence (CPSSA) with intensity-based processing (CLAHE) allows for the generation of high-quality training samples that enhance model generalization without distorting critical biometric features. Future work will focus on improving the computational efficiency of the search process for higher-resolution datasets.
Summary updated