Convolutional Neural Network Approach for Iris Segmentation

Iris segmentation is the initial step for recognition or authentication tasks. In the proposedwork, segmentation of iris region is performed using semantic network. A label or category is associated with every pixel in an image. Semantic segmentation is precise since it clearly detects irregular shaped representations. SegNet is the convolutional neural network proposed for semantic image localization. Pixels with similar attributes are grouped together. Labeled images represent categorical identifiers stored as ground truth masks. Encoder decoder blocks followed by pixel-wise classifier constitute the convolutional network. Each block comprises of convolution, batch normalization and rectified linear unit. Indices determine the mapping of encoder and decoder blocks. Encoder depth regulates the number of times image is upsampled or downsampled. Activations of network relate to the features. In the initial layers, color and edges are learnt. Channels in deeper layers learn the complex features. Learnable parameters are used in convolution. The approach determines iris boundaries without the step of preprocessing. Iris and background are the two labels considered. Eye images and ground truth masks are used for training. Testing samples are evaluated using Jaccard index. The experiment has been conducted on UBIRIS and CASIA datasets for segmentation results, obtaining an F-measure value of 0.987 and 0.962, respectively.