Abstract: Iris recognition is the highly trusted identification recognition technology among methods of biological recognition. In this paper, we use the back propagation algorithm to train the neural network, so as to establish the iris recognition system model. The experiment demonstrates that it has a high recognition rate and the recognition speed is reasonable. The proposed method provides a convenient way for iris recognition.
Introduction
Biometric identification is a cornerstone of modern security, and iris recognition is often cited as one of the most reliable methods due to the iris's unique, stable, and complex texture. However, traditional iris recognition systems often rely on computationally heavy and mathematically complex algorithms—such as Gabor filters, dyadic wavelets, or morphological edge detectors—to extract features before classification. This research explores a simplified pipeline using a multi-layer feed-forward neural network to achieve high accuracy using raw grayscale pixel data as the primary input, thereby reducing the need for specialized feature engineering.
Methodology
The proposed system follows a classic machine learning pipeline consisting of image pre-processing, vectorization, and neural network classification.
Image Pre-processing and Vectorization
Using the OpenCV library, colored iris images are converted to grayscale. To ensure consistency and reduce input dimensionality, images are resized to a standard $100 \times 100$ pixel resolution. This $100 \times 100$ image matrix is then flattened into a $10,000 \times 1$ column vector, where each element represents the grayscale intensity of a single pixel.
Neural Network Architecture
The authors implemented a Back Propagation (BP) neural network with three layers:
- Input Layer: 10,000 neurons, corresponding to the flattened pixel vector.
- Hidden Layer: 40 neurons, determined by the empirical formula $n_2 \le \sqrt{n_1 + n_3} + c$.
- Output Layer: 10 neurons, representing the 10 distinct iris identities used in the experiment.
Training
The network was trained using the BP algorithm with a learning rate ($\eta$) of 0.4 over 5,000 iterations. Weights and thresholds were initialized to small random decimals. The dataset consisted of 10 unique iris images used to establish "learning pairs" for identification.
Results and Discussion
The experimental results indicate that the system is highly effective for small-scale identification tasks.
Recognition Performance
The system achieved a remarkable match rate of 99.946% for the selected objects. The authors noted that the recognition speed is "reasonable," making it potentially viable for real-time applications such as security checkpoints.
Training Stability and Convergence
The study found that the Sum Squared Error (SSE) dropped significantly in the early stages of training. As shown in the error curve, the model reaches peak utility and stabilizes well before the 5,000-iteration mark, with acceptable error levels achieved around 3,000 iterations.
Conclusion
This research demonstrates the effectiveness of using raw grayscale pixel values as features for iris recognition in a BP neural network. While the current study is limited by a small dataset (10 images) and does not account for environmental occlusions like eyelashes or lighting variations, it provides a streamlined foundation for biometric systems. Future work will focus on incorporating structural characteristics to improve robustness and scalability for larger populations.
Summary updated