Convolutional and residual networks for iris recognition

Abstract

Biometrics is essential for authenticating an individual. The existing methods of authentication includes fingerprint scanning, speech recognition, face recognition and iris recognition. The iris recognition is regarded as the most accurate. There are many textural and geometrical elements in a human eye that can be used to uniquely identify an iris pattern. An iris pattern is stable and it is not possible to replicate it. The human iris being a very potential and reliable tool for human identification has the ability to identify individuals with a high degree of assurance. The extraction of good features is the most significant step in an iris recognition system. Initially the iris is augmented. Then, the features can be extracted using a mixed architecture that includes Convolutional Neural Networks(CNN) and residual neural networks. The CNN learns image feature representations automatically. Every neuron accepts input from a small portion of the preceding layer. Weights are made of a set of learnable filters produced randomly and are locally shared. The feature map is the outcome of every filter convolved through the entire image. The pooling layer implements the down sampling operation and decrement the spatial size. The max pooling operation obtains the maximum value from each of a cluster of neurons at the previous layer. The fully connected layer use the extracted features in the preceding layer to do the classification task. The nodes belonging to this layer accepts input from all the nodes in the previous layer. The classifier is needed after feature extraction to find the corresponding label for every test image. The residual network consists of many residual blocks. The presence of an identity mapping distinguishes it from a plain block or convolution block. It has the ability to use knowledge acquired in previous layers. The framework gains advantages from both architectures, i.e., fast convergence from the convolution network and the non-saturation feature from the residual neural network.