Low-resolution periocular images recognition using a novel CNN network

In the context of the COVID-19 epidemic, periocular recognition has become a more effective identity recognition way to replace face recognition. Although various identity recognition model based on convolutional neural networks (CNNs) has achieved satisfactory results, it is limited to high-resolution images. In the previous periocular recognition work, the effect of image resolution is usually ignored. To further extend the application of periocular recognition in unconstrained environments, this work proposes a deep learning-based architecture, AcRNet, to solve the problem of periocular image recognition in different resolutions. AcRNet uses Asymmetric Convolution Block (ACB), and skip connect to enhance the feature extraction capability of the network, and uses a multi-layer feature fusion mechanism to effectively utilize mid-level features. We validated the performance of the proposed network on three datasets, UBIRIS.v2, CASIA_Iris_Distance, and a periocular dataset cropped from VGGface which is called VGGface_periocular, and verified the cross-database performance of the network using a self-acquired periocular dataset. All the low-resolution images were obtained using a bicubic interpolation algorithm. The experimental results demonstrate that AcRNet has a powerful feature extraction capability when extracting features from periocular images of different resolutions. Compared with the previous algorithms, the proposed network has achieved satisfactory performance on the recognition of both high- and low- resolution periocular images. On dataset UBIRIS.v2, AUC values achieved 0.963 and 0.808 when the image resolutions are 128*128 and 8*8.