OcularNet: Deep Patch-based Ocular Biometric Recognition

Abstract- Deep learning played a major role in many recent advancements in mobile ocular biometrics. However, many of the experiments are conducted on models which are large in the number of parameters and are inefficient to deploy on mobile devices. In this paper we propose OcularNet, a convolutions neural network (CNN) model, using patches from the eye images. In OcularNet model, we extract six registered overlapping patches from the ocular and periocular region and train a small convolutions neural network(CNN) for each patch named PatchCNN to extract feature descriptors. As the proposed method is a patch-based technique, one can extract features based on the availability of the region in the eye image. We compare verification performance of the proposed Ocular­Net which has l.SM parameters with the popular ResNet-50 model which has 23.4M parameters. On popular large-scale mobile VISOB dataset, the proposed OcularNet model not only outperformed ResNet-50 with at least 11 % GMR at 1-4 FMR in subject independent verification setting but also has 15.6X less number of parameters. Further, experimental evaluations were performed on UBIRIS-I, UBIRIS-11, and CROSS-EYED datasets to evaluate the performance of Ocular Net over ResNet-50.

I. INTRODUCTION

With the increasing popularity of smartphones, biometrics plays an important role in protecting personal information and data. Several efforts have been made in biometric authen­ tication systems using physical traits such as the face, ocular region, and fingerprint in smartphones. Person recognition using eye regions such as iris, conjunctiva! vasculature, and periocular is known as ocular biometrics [l]. Ocular biometrics at visible spectrum has gained increased interest for smartphones because these images can be easily acquired using the front facing camera of mobile devices bypassing the need of additional hardware for image acquisition. To this front, a large amount of research is conducted in developing robust feature extraction and matching methods for better performance of ocular biometrics in the visible spectrum. Recent large-scale competition on ocular recognition tech­ niques suggests that feature extraction using learning meth­ ods based on neural networks (NN) and convolutional neural network (CNN) outperformed hard-crafted features [2]. In [3], texture features extracted using Maximum Response (MR) filters with deeply coupled autoencoders to extract features from ocular images. In [4], [5], popular models such as VGG[6], ResNet[7], and Inception-net[8] pre-trained on 1.2 rrtillion images from ImageNet database[9] are used to finetune on mobile ocular images using transfer learning for mobile user authentications. In [10], VGG based models are used to train for the periocular recognition without iris and sclear region using data augmentation. In these studies [3], [4], [5], [10], multi-class classifiers are used for recognition and obtained low error rates ( 1 % ). However, both of these methods are a memory and computationally complex to implement on a mobile device. Also, all these models are closed environment classification based methods, where all the subjects present in training are present for verification. To overcome this problem, in this paper we propose a deep learning method to perform patch based ocular biometric recognition. These patches are extracted from six overlap­ping windows of ocular and periocular regions. For each patch, a computationally small deep learning model named PatchCNN is trained to generate feature descriptors. Feature matching is performed by calculating the euclidean distance between each patch of enrollment and verification image pair. Final score for this is calculated using different score fusion techniques such as rrtinimum, mean and median of all scores. Finally, to get the score for input verification image for a given enrollment subject, minimum of all the scores are considered. We evaluate the performance of the proposed OcularNet models in comparison to popular CNN model, ResNet-50[7], on large scale mobile VISOB[2] dataset in open world subject independent verification process where the model is trained on a subset of data which is not used in verification. We also evaluated the performance of the model on datasets with data acquisition methods such as UBIRIS­I [11], UBIRIS-11 [12] and CROSS-EYED [13]. The rest of this paper is organized as follows: Section 2 provides details on proposed OcularNet model. Datasets, experimental protocol, and experimental results are presented in section 3 . Conclusions are drawn in section 4.

With the increasing popularity of smartphones, biometrics plays an important role in protecting personal information and data. Several efforts have been made in biometric authen­ tication systems using physical traits such as the face, ocular region, and fingerprint in smartphones. Person recognition using eye regions such as iris, conjunctiva! vasculature, and periocular is known as ocular biometrics [l]. Ocular biometrics at visible spectrum has gained increased interest for smartphones because these images can be easily acquired using the front facing camera of mobile devices bypassing the need of additional hardware for image acquisition. To this front, a large amount of research is conducted in developing robust feature extraction and matching methods for better performance of ocular biometrics in the visible spectrum. Recent large-scale competition on ocular recognition tech­ niques suggests that feature extraction using learning meth­ ods based on neural networks (NN) and convolutional neural network (CNN) outperformed hard-crafted features [2]. In [3], texture features extracted using Maximum Response (MR) filters with deeply coupled autoencoders to extract features from ocular images. In [4], [5], popular models such as VGG[6], ResNet[7], and Inception-net[8] pre-trained on 1.2 rrtillion images from ImageNet database[9] are used to finetune on mobile ocular images using transfer learning for mobile user authentications. In [10], VGG based models are used to train for the periocular recognition without iris and sclear region using data augmentation. In these studies [3], [4], [5], [10], multi-class classifiers are used for recognition and obtained low error rates ( 1 % ). However, both of these methods are a memory and computationally complex to implement on a mobile device. Also, all these models are closed environment classification based methods, where all the subjects present in training are present for verification. To overcome this problem, in this paper we propose a deep learning method to perform patch based ocular biometric recognition. These patches are extracted from six overlap­ping windows of ocular and periocular regions. For each patch, a computationally small deep learning model named PatchCNN is trained to generate feature descriptors. Feature matching is performed by calculating the euclidean distance between each patch of enrollment and verification image pair. Final score for this is calculated using different score fusion techniques such as rrtinimum, mean and median of all scores. Finally, to get the score for input verification image for a given enrollment subject, minimum of all the scores are considered.

We evaluate the performance of the proposed OcularNet models in comparison to popular CNN model, ResNet-50[7], on large scale mobile VISOB[2] dataset in open world subject independent verification process where the model is trained on a subset of data which is not used in verification. We also evaluated the performance of the model on datasets with data acquisition methods such as UBIRIS­I [11], UBIRIS-11 [12] and CROSS-EYED [13]. The rest of this paper is organized as follows: Section 2 provides details on proposed OcularNet model. Datasets, experimental protocol, and experimental results are presented in section 3 . Conclusions are drawn in section 4.