sondosaabed/iris-of-eyes-recognition

👁️ A Biometric Authentication system using Iris. Enrollment and Authentication Modules. End to End, Iris Segmentation Free using DCNNs, Accuracy of 93.15% 👁️

🔴 Cite This Notebook 🔴

If you find this notebook useful in your research or projects, please consider citing it. Proper citation helps me gain recognition for my work and allows others to follow and build upon it.

Sondos Aabed, An End-to-end segmentation-free approach Iris Biometric Authentication, Open Source Contribution (GitHub & Kaggle), May 2024. Available at: https://github.com/sondosaabed/Iris-of-eyes-recognition and https://www.kaggle.com/code/sondosaabed/iris-eye-recognition-endtoend-93

---

Iris-of-eyes-recognition

In this project, a Biometric Authentication system using the Iris biometric authentication method is designed. The approach taken is using the CASIA-Thousand-IRIS dataset and model it using the Deep Convultional Neural Network Architicture, with the minimum image-preprocessing such as resizing with keeping the aspect ratiio and normalization. It is an end-to-end technique without performing segmentaion of the IRIS itself. The results are promising, even without perfroing training on augmentation, the testing accuracy has reached (91.10%). Finally, for the proof of the (biometric authentication system concept) a simple mobile application is designed and the model is deployed on it (IrisRecognizer) as it was exported to it's liter version were default quantization is performed.

Model Training (Accuracy and Loss Learning Curves)

Example of results on testing dataset

GUI Authenticater Module

In a biometric authentication system there has to be an authenticter part for it. A simple mobile application is done for the concept:

Refer to this repository: https://github.com/sondosaabed/IrisRecognizer

iris-of-eyes-recognition's Issues

Model Export and Test Doing Inference

• Export to tflite
• Test exported class
• Deploy the model on the mobile

Fluctuating Validation Performance

• Work on stabilizing training before the 20th epoch, the validation is fluctuating

Multiclass Metrics of Evaluation

• False Match Rate (FMR): The rate at which an imposter is incorrectly classified as a genuine user.
• False Non-Match Rate (FNMR): The rate at which a genuine user is incorrectly classified as an imposter.
• Plot the Receiver Operating Characteristic (ROC) curve, which shows the trade-off between FMR and FNMR.
• Determine the Equal Error Rate (EER), the point on the ROC curve where FMR and FNMR are equal.