Basically, boundary-based iris segmentation methods require prominent contrast of structure components. Gradient and contour information was concerned more in such methods. While pixelbased methods rely highly on discriminative features such asimages’ texture, color and intensity. Also, approaches such as [30,31] integrated these two kind of methods. [30] first roughly cluster image pixels into iris and non-iris regions by setting threshold on brightness, and then on the obtained coarse iris location image, an integro-differential model was adopted to locate iris boundary. While [31] first used Random Walker to locate coarse boundary circle of iris region, then a series of operations based on statistical gray level intensity information were adopted for pixel-level boundary refine.Deep learning based methods. Benefited from large-scale data collection, rapid development on computing performance and fastGPU implementations of artificial neural networks [32–35], since 2010s, deep learning-based method dramatically boost in the field of computer vision, as well as the field of image segmentation [36–39]. Unlike traditional patch classification-based CNN models that using fully connected layers after convolutional layers to get fixed length feature vectors, FCN [39] allow arbitrary input image size and adopt deconvolution layer for upsampling the different convolutional layers’ feature maps to target size. Compared with former approaches, FCN avoided separately running network for each patch, and boosted segmentation speed. Papers including [11,12]introduce modified FCN to the task of iris segmentation. However,because the feature maps in FCN for upsampling is too coarse,FCN’s segmentation results are not fine enough. Unlike FCN that upsampling different size coarse feature maps to target resolution,U-Net [13] reform a lot in upsampling stage. U-Net adopt an encoder-decoder structure, and in U-Net’s successive layers, pooling operators are replaced by a serial of upsampling operators,which makes the whole network a symmetrical U-shaped model.U-Net has been proven good performance in the field of biomedical[40–42], and can work with relatively few training images and yields more precise segmentations.