IRIS RECOGNITION ARCHITECTURE WITH CONVOLUTIONAL NEURAL NETWORK FOR NON-IDEAL IMAGES ON VISIBLE SPECTRUM

• Main Author: Prabowo, Harry
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/56335
• Institution: Institut Teknologi Bandung

Iris has not seen widespread application as a biometric technique, however superior it is to competitors. Ideal imaging is one of the key challenges preventing such application. Recent solutions rely on near-infrared spectrum – a more resilient alternative given non-ideal images. Solutions as such requires special hardware, hence suppressing adoption. This paper proposes an iris recognition architecture for non-ideal imaging on visible spectrum using convolutional neural network (CNN) due to the semantic biometric features of iris. This paper hypothesizes that superresolution and better segmentation help iris recognition for non-ideal images on visible spectrum, The architecture consists of three modules: (1) segmentation with Mask-RCNN; (2) super-resolution with EDSR; and (3) classification with DenseNet201 as a control module. For comparison, the classic Circle Hough Transform (CHT) for iris segmentation will be used together with the control module as baseline. This paper also explores the relationship between said techniques and recognition performance, while pursuing an effective iris recognition solution on non-ideal visible spectrum imaging. Tests and evaluations are performed with UBIRIS.v2 dataset on platforms designed for end users. Segmentation by cropping irises using Mask-RCNN’s bounding box yields an improvement in accuracy by 1.62 times – averaging at 84.43% in 0.75 seconds. This is further improved by EDSR, increasing the accuracy further by 1.04% while taking 1.47 seconds. Meanwhile, instance segmentation using Mask-RCNN yields an improvement of 1.13 times – averaging at 68.51%, which is further improved by EDSR to 69.72%.