Multitype Convolutional Networks for Improved Iris Classification

Iris recognition is an essential biometric modality that has achieved far-reaching adoption in most applications including government ID plans, voter registration and de-duplication, international border crossing, and mobile device authentication. Due to its high sensitivity as a secure, non-invasive human identification technique, iris recognition has various challenges like light variation, specular highlights, motion blur, and image noise to deal with. To overcome these limitations, deep learning models have been employed more widely both for pattern matching and feature extraction so that higher quality and robust detection of complex and highly unique iris textures can be achieved. The current paper presents a successful two-stage iris recognition system. Iris segmentation and detection are achieved in the first stage by using the HoughCircles algorithm and inpainting improved to achieve higher image quality and remove artifacts. The second stage employs a new deep network employing multiple types of convolutional layers like regular, spatial, and depthwise convolutions to execute efficient feature extraction and accurate classification. The proposed model was also validated and tested with the MULBv1 iris image database and achieved promising results with the classification accuracy of 98.21% and F1-score of 98.16%. The results prove the effectiveness of the proposed method as a feasible and practical solution to be used in real-life iris identification systems.