Ocular Biometrics by Score-Level Fusion of Disparate ExpertsOcular Biometrics by Score-Level Fusion of Disparate Experts
The concept of periocular biometrics has been gaining relevance, in particular to improve the robustness of iris recognition to degraded data. This paper proposes an atomistic periocular recognition algorithm, in the sense that describes a recognition ensemble made of two disparate components, with radically different properties: the best expert analyses the iris texture and exhaustively exploits the multi-spectral information in visible-light data; complementary, another expert parameterises the shape of eyelids and defines a surrounding dimensionless region-of-interest, from where statistics of the eyelids, eyelashes and skin wrinkles/furrows are encoded. Both experts work on disjoint data and use very different encoding/matching strategies, meeting three important properties: 1) experts produce practically independent responses, which is behind the better performance of the ensemble when compared to the best individual recogniser; 2) experts are not particularly sensitive to the same image covariate, which accounts for augmenting the robustness against degraded data; and 3) experts disregard information in the periocular region that can be easily forged (e.g., shape of eyebrows), which can the regarded as an active anti-counterfeit measure. An empirical evaluation was conducted on two public data sets (FRGC and UBIRIS.v2), and points for consistent improvements in performance of the proposed ensemble over the state-of-the-art periocular recognition algorithms.
Vantage-Point: A Real-Time Classification Strategy
This paper formalizes a data structure and an algorithm for feature classification. In learning, a data structure that resembles the idea of Vantage-Point tree recursively divides the feature space into disjoint subspaces. At each subset, the misclassification rate attained by linear discriminant analysis determines further divisions on that space, ending up with an ensemble of classifiers at the tree leaves. In classification, only a reduced number of classifiers are considered, in a procedure that runs in time roughly logarithmic with respect to the size of the learning set, making it suitable for real-time data processing. The proposed method was empirically validated in widely known and freely available data sets, attaining misclassification rates that are favorably comparable to state-of-the-art individual and ensemble classifiers.
A knowledge-based approach to the iris segmentation problem
This paper describes a knowledge-based approach to the problem of locating and segmenting the iris in images showing close-up human eyes. This approach is inspired in the expert system's paradigm but, due the specific processing problems associated with image analysis, uses direct encoding of the "decision rules", instead of a classic, formalized, knowledge base. The global approach used in this paper can be useful to solve other image analysis problems over which human "experts" have better performance than the present computer-based solutions.
DropAll: Generalization of Two Convolutional Neural Network Regularization Method
We introduce DropAll, a generalization of DropOut and DropConnect, for regularization of fully-connected layers within convolutional neural networks. Applying these methods amounts to sub-sampling a neural network by dropping units. Training with DropOut, a randomly selected subset of activations are dropped, when training with DropConnect we drop a randomly subsets of weights. With DropAll we can perform both methods. We show the validity of our proposal by improving the classification error of networks trained with DropOut and DropConnect, on a common image classification dataset. To improve the classification, we also used a new method for combining networks.
3D Object Recognition using Convolutional Neural Networks with Transfer Learning between Input Channels
RGB-D data is getting ever more interest from the research community as both cheap cameras appear in the market and the applications of this type of data become more common. A current trend in processing image data is the use of convolutional neural networks (CNNs) that have consistently beat competition in most benchmark data sets. In this paper we investigate the possibility of transferring knowledge between CNNs when processing RGB-D data with the goal of both improving accuracy and reducing training time. We present experiments that show that our proposed approach can achieve both these goals.
On Indexing and Retrieving Degraded Iris Biometric Signatures
Recent advances in the deployment of biometric systems at large scale identification scenarios augmented the interest in indexing/retrieval strategies. In this paper it is proposed a technique that operates at the IrisCode level. Gallery codes are decomposed at multiple scales and, based on the most reliable components of each one, the position of the corresponding identity in nodes of a n-ary tree determined. In retrieval, the probe is similarly decomposed and the distance to multi-scale centroids used to privilege paths on the tree. This way, for each query only a small number of branches are traversed up to the last level, which nodes most times contain the identity of interest. When compared with indexing strategies that operate at the code level, the proposed method behaves comparably on good-quality data and shows higher robustness when handling degraded codes, Lastly, the temporal computational requirements of the proposed strategy are discussed and compared with related techniques.