Deep Learning Based Off-Angle Iris Recognition

This paper investigates the effect of segmentation errors on off-angle iris recognition. We first segment the inner and outer boundaries of off-angle iris images by fitting the best ellipses to boundaries. Then, we validate the segmentation results using Ground Truth Tool and fix the possible errors. Second, we add random errors on iris segmentation parameters to examine how Hamming distance distribution changes for different amount of segmentation error. In order to add errors, we first group the parameters into three sets as ellipse center (i.e., x, y), minor and major axis (i.e., r 1 , r 2 ) and orientation (θ) and changed the parameters by adding positive and negative random error as a noise to the ground truth segmentation parameters. Our purpose is to show how performance of iris recognition is affected if there is an error in segmentation and how important is to have a very robust segmentation algorithm for off-angle iris recognition. We test the effect of segmentation errors using our off-angle iris dataset that contains images from -50° to +50° in gaze angles. Based on our experimental results, we observed that error in ellipse center decrease the recognition performance worse than minor and major axis and orientation. The main reason is the normalization process tolerates the small amount of error in minor and major axis and bit shift in matching handles the small variations in the orientation errors. However, if the error in the ellipse orientation exceeds the limit of the bit shift method tolerance, the performance of the off-angle iris recognition dramatically drops.

Standoff iris recognition systems have been developed to identify not only fully cooperative individuals but also non-cooperative moving subjects. Since iris images are captured in much less constrained conditions, they are likely to be non-ideal including being off-angle. However, traditional iris recognition systems are designed for only frontal iris images. Therefore, the existing databases save binary iris codes using only frontal images for iris matching. To use these existing iris codes, standoff systems need to be compatible with existing iris matching algorithms by generating the frontal version of iris codes even for off-angle images. Since gaze estimation is the first essential component to reconstruct the frontal view of off-angle iris images, we first need to estimate the eye gaze direction from a single off-angle iris image. This paper presents a deep learning-based gaze angle estimation method for standoff iris recognition frameworks. As the main contribution, our approach will allow us to estimate the gaze angle from an off-angle image using transfer learning and a regression model without segmentation and using additional hardware. The proposed CNN-based gaze estimation method is trained and tested with 108,600 off-angle images from 100 subjects. It shows high accuracy in gaze estimation with an average error of 4° in angle for off-angle images captured from every angle ranging -50° to +50° even if training images are not available for all angles.

The ability to identify individuals based on their iris is known as iris recognition. Over the past decade iris recognition has garnered much attention because of its strong performance in comparison with other mainstream biometrics such as fingerprint and face recognition. Performance of iris recognition systems is driven by application scenario requirements. Standoff distance, subject cooperation, underlying optics, and illumination are a few examples of these requirements which dictate the nature of images an iris recognition system has to process. Traditional iris recognition systems, dubbed "stop and stare", operate under highly constrained conditions. This ensures that the captured image is of sufficient quality so that the success of subsequent processing stages, segmentation, encoding, and matching are not compromised. When acquisition constraints are relaxed, such as for surveillance or iris on the move, the fidelity of subsequent processing steps lessens.;In this dissertation we propose a multi-faceted framework for mitigating the difficulties associated with non-ideal iris. We develop and investigate a comprehensive iris image quality metric that is predictive of iris matching performance. The metric is composed of photometric measures such as defocus, motion blur, and illumination, but also contains domain specific measures such as occlusion, and gaze angle. These measures are then combined through a fusion rule based on Dempster-Shafer theory. Related to iris segmentation, which is arguably one of the most important tasks in iris recognition, we develop metrics which are used to evaluate the precision of the pupil and iris boundaries. Furthermore, we illustrate three methods which take advantage of the proposed segmentation metrics for rectifying incorrect segmentation boundaries. Finally, we look at the issue of iris image interoperability and demonstrate that techniques from the field of hardware fingerprinting can be utilized to improve iris matching performance when images captured from distinct sensors are involved.

Iris recognition is among the highest accuracy biometrics. However, its accuracy relies on controlled high quality capture data and is negatively affected by several factors such as angle, occlusion, and dilation. Non-ideal iris recognition is a new research focus in biometrics. In this paper, we present a gaze estimation method designed for use in an off-angle iris recognition framework based on the ORNL biometric eye model. Gaze estimation is an important prerequisite step to correct an off-angle iris images. To achieve the accurate frontal reconstruction of an off-angle iris image, we first need to estimate the eye gaze direction from elliptical features of an iris image. Typically additional information such as well-controlled light sources, head mounted equipment, and multiple cameras are not available. Our approach utilizes only the iris and pupil boundary segmentation allowing it to be applicable to all iris capture hardware. We compare the boundaries with a look-up-table generated by using our biologically inspired biometric eye model and find the closest feature point in the look-up-table to estimate the gaze. Based on the results from real images, the proposed method shows effectiveness in gaze estimation accuracy for our biometric eye model with an average error of approximately 3.5 degrees over a 50 degree range.

In recent years, the topic of contactless biometric identification has gained considerable traction due to the COVID-19 pandemic. One of the most well-known identification technologies is iris recognition. Determining the classification threshold for large datasets of iris images remains challenging. To solve this issue, it is essential to extract more discriminatory features from iris images. Choosing the appropriate loss function to enhance discrimination power is one of the most significant factors in deep learning networks. This paper proposes a novel iris identification framework that integrates the light-weight MobileNet architecture with customized ArcFace and Triplet loss functions. By combining two loss functions, it is possible to improve the compactness within a class and the discrepancies between classes. To reduce the amount of preprocessing, the normalization step is omitted and segmented iris images are used directly. In contrast to the original SoftMax loss, the EER for the combined loss from ArcFace and Triplet is decreased from 1.11% to 0.45%, and the TPR is increased from 99.77% to 100%. In CASIA-Iris-Thousand, EER decreased from 4.8% to 1.87%, while TPR improved from 97.42% to 99.66%. Experiments have demonstrated that the proposed approach with customized loss using ArcFace and Triplet can significantly improve state-of-the-art and achieve outstanding results.

It has been proven that hamming distance score between frontal and off-angle iris images of same eye differs in iris recognition system. The distinction of hamming distance score is caused by many factors such as image acquisition angle, occlusion, pupil dilation, and limbus effect. In this paper, we first study the effect of the angle variations between iris plane and the image acquisition systems. We present how hamming distance changes for different off-angle iris images even if they are coming from the same iris. We observe that increment in acquisition angle of compared iris images causes the increment in hamming distance. Second, we propose a new technique in off-angle iris recognition system that includes creating a gallery of different off-angle iris images (such as, 0, 10, 20, 30, 40, and 50 degrees) and comparing each probe image with these gallery images. We will show the accuracy of the gallery approach for off-angle iris recognition.

The human iris is traditionally imaged in Near Infrared (NIR) wavelengths (700nm-900nm) for iris recognition. The absorption co-efficient of color inducing pigment in iris, called Melanin, decreases after 700nm thus minimizing its effect when iris is imaged at wavelengths greater than 700nm. This thesis provides an overview and explores the efficacy of iris recognition at different wavelength bands ranging from visible spectrum (450nm-700nm) to NIR (700nm-900nm) and Short Wave Infrared (900nm-1600nm). Different matching methods are investigated at different wavelength bands to facilitate cross-spectral iris recognition.;The iris recognition analysis in visible wavelengths provides a baseline performance when iris is captured using common digital cameras. A novel blob-based matching algorithm is proposed to match RGB (visible spectrum) iris images. This technique generates a match score based on the similarity between blob like structures in the iris images. The matching performance of the blob based matching method is compared against that of classical 'Iris Code' matching method, SIFT-based matching method and simple correlation matching, and results indicate that the blob-based matching method performs reasonably well. Additional experiments on the datasets show that the iris images can be matched with higher confidence for light colored irides than dark colored irides in the visible spectrum.;As part of the analysis in the NIR spectrum, iris images captured in visible spectrum are matched against those captured in the NIR spectrum. Experimental results on the WVU multispectral dataset show promise in achieving a good recognition performance when the images are captured using the same sensor under the same illumination conditions and at the same resolution. A new proprietary 'FaceIris' dataset is used to investigate the ability to match iris images from a high resolution face image in visible spectrum against an iris image acquired in NIR spectrum. Matching in 'FaceIris' dataset presents a scenario where the two images to be matched are obtained by different sensors at different wavelengths, at different ambient illumination and at different resolution. Cross-spectral matching on the 'FaceIris' dataset presented a challenge to achieve good performance. Also, the effect of the choice of the radial and angular parameters of the normalized iris image on matching performance is presented. The experiments on WVU multispectral dataset resulted in good separation between genuine and impostor score distributions for cross-spectral matching which indicates that iris images in obtained in visible spectrum can be successfully matched against NIR iris images using 'IrisCode' method.;Iris is also analyzed in the Short

The traditional iris recognition systems require equal high quality human iris images. A cheap image acquisition system has difficulty in capturing equal high quality iris images. This paper describes a new feature representation method for iris recognition robust to noises. The disc-shaped iris image is first convolved with a low pass filter along the radial direction. Then, the radially smoothed iris image is decomposed in the angular direction using a one-dimensional continuous wavelet transform. Each decomposed one-dimensional waveform is approximated by an optimal piecewise linear curve connecting a small set of node points. The set of node points is used as a feature vector. The optimal approximation procedure reduces the feature vector size while maintaining recognition accuracy. The similarity between two iris images is measured by the normalized cross-correlation coefficients between optimal curves. The similarity between two iris images is estimated using mid-frequency bands. The rotation of one-dimensional signals due to the head tilt is estimated using the lowest frequency component. Experimentally we show the proposed method produces superb performance in iris recognition.

The Iris recognition is commonly used as a security system due to its robustness against imposters. Iris datasets are huge and hence those datasets occupy more space. Iris image compression has become an important part of better performance like speed and data storage. The portable Iris system is in huge demand. That portable systems need to transmit the iris images through a very small bandwidth channel. To reduce the time for transferring a huge number of data over small bandwidth channel, iris file can be compressed to some extent to minimize the size. Another problem is that when an image is captured, it captures some noise that disturbs the recognition performance, so, denoising is required for noise-free images. This paper separately analyzes the impact of wavelet compression along with denoising on iris images. The compression analysis is done using Embedded Zero Tree Wavelet, the other technique used is Set Partitioning in Hierarchical Tree and the third technique used is Spatial-Orientation Tree Wavelet. Denoising is done using different wavelets Daubechies, Haar, Biorthogonal and Fejer-Korovkin. The impact of the wavelet compression and denoising techniques on recognition performance are compared with False Rejection Rate and False Acceptance Rate. The quality of a compressed image is calculated with different quality metrics. This work establishes that compression and denoising of the images minimally affect the recognition performance.

The popularity of the iris biometric has grown considerably over the past two to three years. Most research has been focused on the development of new iris processing and recognition algorithms for frontal view iris images. However, a few challenging directions in iris research have been identified, including processing of a nonideal iris and iris at a distance. In this paper, we describe two nonideal iris recognition systems and analyze their performance. The word "nonideal" is used in the sense of compensating for off-angle occluded iris images. The system is designed to process nonideal iris images in two steps: 1) compensation for off-angle gaze direction and 2) processing and encoding of the rotated iris image. Two approaches are presented to account for angular variations in the iris images. In the first approach, we use Daugman's integrodifferential operator as an objective function to estimate the gaze direction. After the angle is estimated, the off-angle iris image undergoes geometric transformations involving the estimated angle and is further processed as if it were a frontal view image. The encoding technique developed for a frontal image is based on the application of the global independent component analysis. The second approach uses an angular deformation calibration model. The angular deformations are modeled, and calibration parameters are calculated. The proposed method consists of a closed-form solution, followed by an iterative optimization procedure. The images are projected on the plane closest to the base calibrated plane. Biorthogonal wavelets are used for encoding to perform iris recognition. We use a special dataset of the off-angle iris images to quantify the performance of the designed systems. A series of receiver operating characteristics demonstrate various effects on the performance of the nonideal-iris-based recognition system.

Plant fungal diseases are one of the most important causes of crop yield losses. Therefore, plant disease identification algorithms have been seen as a useful tool to detect them at early stages to mitigate their effects. Although deep-learning based algorithms can achieve high detection accuracies, they require large and manually annotated image datasets that is not always accessible, specially for rare and new diseases. This study focuses on the development of a plant disease detection algorithm and strategy requiring few plant images (Few-shot learning algorithm). We extend previous work by using a novel challenging dataset containing more than 100,000 images. This dataset includes images of leaves, panicles and stems of five different crops (barley, corn, rape seed, rice, and wheat) for a total of 17 different diseases, where each disease is shown at different disease stages. In this study, we propose a deep metric learning based method to extract latent space representations from plant diseases with just few images by means of a Siamese network and triplet loss function. This enhances previous methods that require a support dataset containing a high number of annotated images to perform metric learning and few-shot classification. The proposed method was compared over a traditional network that was trained with the cross-entropy loss function. Exhaustive experiments have been performed for validating and measuring the benefits of metric learning techniques over classical methods. Results show that the features extracted by the metric learning based approach present better discriminative and clustering properties. Davis-Bouldin index and Silhouette score values have shown that triplet loss network improves the clustering properties with respect to the categorical-cross entropy loss. Overall, triplet loss approach improves the DB index value by 22.7% and Silhouette score value by 166.7% compared to the categorical cross-entropy loss model. Moreover, the F-score parameter obtained from the Siamese network with the triplet loss performs better than classical approaches when there are few images for training, obtaining a 6% improvement in the F-score mean value. Siamese networks with triplet loss have improved the ability to learn different plant diseases using few images of each class. These networks based on metric learning techniques improve clustering and classification results over traditional categorical cross-entropy loss networks for plant disease identification.

In order to improve the performance of iris recognition, a novel method for iris recognition based on block theory and self-adaptive feature selection is proposed in this paper. Firstly, the normalized iris image is decomposed by convolving with multi-scale and multi-orientation Gabor filters, and then separated into several blocks, the block feature vector which includes mean and variance of Gabor coefficients inside each block can be obtained through statistical techniques, the iris feature vector of the whole iris image is then constructed by conjugating the block feature vector in row column order, finally the two-classifier of iris image are established based on the most distinguishable features, and the multi-classifiers of iris image are established by voting mechanism, and the performance is test by CASIA iris database. The results show that, compared with the traditional iris recognition methods, the proposed method has improved the iris recognition rate.

TripletMultiDTI: Multimodal representation learning in drug-target interaction prediction with triplet loss function

Copper elbows are an important product in industry. They are used to connect pipes for transferring gas, oil, and liquids. Defective copper elbows can lead to serious industrial accidents. In this paper, a novel model named YOT-Net (YOLOv3 combined triplet loss network) is proposed to automatically detect defective copper elbows. To increase the defect detection accuracy, triplet loss function is employed in YOT-Net. The triplet loss function is introduced into the loss module of YOT-Net, which utilizes image similarity to enhance feature extraction ability. The proposed method of YOT-Net shows outstanding performance in copper elbow surface defect detection.

The role of biometrics, as an integral component of security infrastructures, has been firmly established in today's world. Its ability to accurately identify people from a distance has served as an important life-saving tool in the arsenalof our men and women in uniform, out in war theaters; it has also served as an effective forensics tool for our police force and off-late, is being increasingly adopted to secure sensitive transactions such as those with financial institutions.Traditionally, fingerprint patterns and facial photographs have held sway over identity veri fication tools; the former, because of its uniqueness and the latter because of its ubiquity. However, over the past decade, iris patterns have alsocome to the fore because of the combined advantage of uniqueness across the populace as well as the ability to image these patterns from a distance. Today most commercial iris acquisition systems report recognition accuracies upwards of 90%, over entire populations. However, the caveat is that these systems areconstrained to operate very close to the user, owing to the challenges involved in achieving sufficient resolution over the iride texture from large distances. This dissertation aims to push the boundaries of iris imaging devices and in-troduces theoretical and practical considerations in designing a novel long range iris recognition system. The aim is to design a system capable of acquiring enrollment quality iris images, similar in quality to those captured by short range systems such as IriShield, LG, PIER, HIIDE systems, at much larger stand-off distances. The challenges involved in acquiring images with sufficient spatialresolution, when subjects are too far out from the optimal acquisition range, are explored and potential solutions are described. First, we describe the design methodology behind creating a long range iris system, to aid the reader inunderstanding the key design considerations when using available commercial o -the-shelf hardware. Competing objectives such as achieving high magnifi cation while maintaining sufficient depth of field, using a sensor with large pixel pitch for low noise performance while resolving the required patterns, etc. are discussed. As a result of this work, the rst long range iris acquisition system of its kind, capable of acquiring enrollment iris images from up to 9mand matching up to 12m, has been designed and built. We show the system, present lessons learned during the design process as well as empirical results for the imaging performance. The latter half of the dissertation develops aniris texture enhancement algorithm for restoring degraded iris images - specifically noisy, low-resolution and out-of-focus/blurred images, which are often encountered in the real world. The method is developed and evaluated usingsynthesized degraded images from the NIST Iris Challenge Evaluation (ICE) database. A signifi cant improvement in verification performance is also seen, when using this enhancement method, on actual degraded images captured with the proposed long range iris system. The proposed enhancement method is modeled as a linear regression problem that exploits the sparsity of iris textures inan over-complete dictionary. Improvements in iris recognition performance using the enhanced textures are studied in the latter half of this dissertation. In summary, this work enhances and exceeds the capabilities of current iris imaging systems for the explicit purpose of accurately identifying threats using iride patterns, from distances larger than are currently possible. This will serve as anecessary tool in the arsenal of security establishments to identify threats, save lives and to protect our freedoms for generations to come.