Source-linked AI summary
An Experimental Study of Deep Convolutional Features For Iris Recognition
Shervin Minaee, Amirali Abdolrashidi, Yao Wang
TL;DR
Iris recognition traditionally depends on hand-crafted features and preprocessing, while the transferability of deep features to distinguishing people within one object category remained to be evaluated. This paper extracts VGG-Net features without fine-tuning, reduces them with PCA, and classifies them using multi-class SVM across two iris datasets. The approach achieves promising results, including a best accuracy of 99.4%, and outperforms the previous best result on one database.
Problem
Traditional iris recognition relies on hand-crafted features and substantial preprocessing, motivating evaluation of whether object-trained deep features transfer to distinguishing subjects within the iris category.
Method
The study extracts features from VGG-Net without fine-tuning, applies PCA for dimensionality reduction, and uses multi-class SVM for recognition.
Results
99.4% best accuracy was achieved, with the method outperforming the previous best result on one database.
Takeaways & Limitations
VGG-Net features trained for object recognition can transfer to biometric recognition across two tested iris datasets without preprocessing or architecture optimization.
Abstract
from arXiv · showhide
Iris is one of the popular biometrics that is widely used for identity authentication. Different features have been used to perform iris recognition in the past. Most of them are based on hand-crafted features designed by biometrics experts. Due to tremendous success of deep learning in computer vision problems, there has been a lot of interest in applying features learned by convolutional neural networks on general image recognition to other tasks such as segmentation, face recognition, and object detection. In this paper, we have investigated the application of deep features extracted from VGG-Net for iris recognition. The proposed scheme has been tested on two well-known iris databases, and has shown promising results with the best accuracy rate of 99.4\%, which outperforms the previous best result.
1. INTRODUCTION
Iris recognition has traditionally relied on hand-crafted features and substantial preprocessing, motivating evaluation of transferable deep features. The paper applies VGG-Net features without fine-tuning or iris segmentation and reports strong recognition performance despite dataset variation.
- Motivation: Traditional iris systems typically segment and normalize the iris before extracting hand-crafted features tuned to particular data.Their performance on other biometrics or datasets is not guaranteed.
- Motivation: Deep neural networks offer learned features that can transfer across tasks, reducing dependence on task-specific preprocessing and parameter tuning.
- Approach: The study extracts VGG-Net features without fine-tuning, then applies PCA for dimensionality reduction and multi-class SVM for recognition.Experiments use the CASIA 1000 Iris and IIT iris datasets.
- Findings: The framework skips iris segmentation to test robustness to intra-class variation and still achieves a very high accuracy rate on CASIA-1000.
- Findings: VGG-Net features work for distinguishing subjects within the same object category, although the network was trained for object classification.
2. FEATURES
The paper uses VGG-Net as a learned multi-layer feature extractor and contrasts it with predefined-filter scattering representations. VGG-Net provides a homogeneous deep architecture whose features are evaluated across layers and reduced with PCA.
- Deep features: VGG-Net learns deep image features directly, unlike hand-crafted descriptors designed by computer vision experts.Convolutional networks have achieved strong results across computer vision benchmarks.
- Transfer to iris recognition: The study investigates whether VGG-Net features transfer from object recognition to iris recognition, which distinguishes people rather than object categories.
- VGG-Net architecture: VGG-Net uses 16 convolutional or fully connected layers, five pooling layers, and approximately 138 million parameters.Its architecture uses homogeneous 3x3 convolutions and 2x2 pooling operations.
- Feature evaluation: Features are extracted from multiple VGG-Net layers and evaluated for iris recognition, with PCA used to reduce their dimensionality.
- Alternative representation: Scattering networks use predefined filters across multiple scales and orientations, then concatenate statistical features from transformed images.Their second-layer outputs illustrate orientation- and scale-specific edge information.
3. RECOGNITION ALGORITHM
Recognition uses multi-class SVM after feature extraction. The classifier selects separating hyperplanes with maximum margin and extends binary classification to multiple classes.
- Classification: After feature extraction, a classifier assigns each test image to a corresponding label.
- Binary SVM: For linearly separable binary data, SVM finds a hyperplane with the maximum margin between classes.The hyperplane is represented as w.x+b = 0.
- Binary SVM: The SVM training problem is constrained by yi(w.xi + b) ≥1 for i = 1, . . . , n.
- Decision function: The resulting classifier uses learned αi values and bias b to compute the sign of a weighted decision function.
- Multi-class SVM: For M classes, one-vs-all SVM trains M binary classifiers and assigns the class with the largest margin.A one-versus-one alternative selects the class receiving the majority of classifier votes.
4. EXPERIMENTAL RESULTS AND ANALYSIS
The study evaluates VGG-Net features for iris recognition across two databases, varying feature layers, PCA dimensionality, and training-sample counts. The method achieves high recognition accuracy, including above 98% with few PCA features and features from layers after the seventh.
- Experimental setup: The experiments use VGG-Net features, PCA-based dimensionality reduction, and a linear SVM on CASIA-Iris-1000 and IIT Delhi databases.Features are extracted from multiple VGG-Net layers, with average filter outputs forming layer-specific vectors.
- PCA dimensionality: Above 98% accuracy is achieved on IIT Delhi using 100 PCA features, with only about a 1% increase when more features are used.The result is reported for fc6 features after PCA.
- VGG-Net layers: Above 98% accuracy is achieved with features from any VGG-Net layer after the seventh on IIT Delhi.Accuracy peaks at the tenth layer and then decreases.
- Training samples: Recognition accuracy gains substantially when using 3 of 10 training samples per person and remains relatively constant with additional samples.The study varies the number of training samples from 1 to 5 per person on the IIT database.
- Comparison: The proposed deep-feature scheme achieves the highest accuracy on the IIT database among the compared algorithms and does not require iris segmentation.The authors attribute the performance to the richness of deep features, while noting segmentation could improve difficult cases.
5. CONCLUSION
The conclusion evaluates VGG-Net deep features with a simple classifier for iris recognition. Across two datasets, the features transfer from object recognition to biometric recognition, outperforming previous results on one database without preprocessing or architecture optimization.
- The paper evaluates VGG-Net deep features followed by a simple classification algorithm for iris recognition.
- Features learned for object recognition transfer effectively to biometric recognition across two well-known iris datasets.
- The method achieves promising results and outperforms previous best results on one of the databases without preprocessing or architecture optimization.