The high demand for ever more accurate biometric systems has driven the search for methods that reconstruct the iris surface in a 3D model. The intent in adding the depth dimension is to improve accuracy even in large databases. Here, we present a novel approach to iris recognition from 3D models. First, the iris 3D model is reconstructed from a single image using irisDepth, a CNN based method. Then, a 3D descriptor called Spin Image is obtained for keypoints of the 3D model. After that, matches are found between keypoints in the query and the reference 3D models using k-dimensional trees. Finally, those keypoint matches are used to determine the spatial transformation that best aligns the 3D models. A combination of the transformation error and the inlier ratio is used as the metric to assess the similarity of two iris 3D models. We applied this method in a dataset of 100 eyes and 2,000 iris 3D models. Our results indicate that using the proposed method is more effective than alternative methods, such as Dougman's iris code, point-to-point distance between the 3D models, the 3D rubber sheet model, and CNN-based methods.

In most iris recognition systems the texture of the iris image is either the result of the interaction between the iris and Near Infrared (NIR) light, or between the iris pigmentation and visible-light. The iris, however, is a three-dimensional organ, and the information contained on its relief is not being exploited completely. In this article, we present an image acquisition method that enhances viewing the structural information of the iris. Our method consists of adding lateral illumination to the visible light frontal illumination to capture the structural information of the muscle fibers of the iris on the resulting image. These resulting images contain highly textured patterns of the iris. To test our method, we collected a database of 1,920 iris images using both a conventional NIR device, and a custom-made device that illuminates the eye in lateral and frontal angles with visible-light (LFVL). Then, we compared the iris recognition performance of both devices by means of a Hamming distance distribution analysis among the corresponding binary iris codes. The ROC curves show that our method produced more separable distributions than those of the NIR device, and much better distribution than using frontal visible-light alone. Eliminating errors produced by images captured with different iris dilation (13 cases), the NIR produced inter-class and intra-class distributions that are completely separable as in the case of LFVL. This acquisition method could also be useful for 3D iris scanning.

Trademarks are recognizable images and/or words used to distinguish various products or services. They become associated with the reputation, innovation, quality, and warranty of the products. Countries around the world have offices for industrial/intellectual property (IP) registration. A new trademark image in application for registration should be distinct from all the registered trademarks. Due to the volume of trademark registration applications and the size of the databases containing existing trademarks, it is impossible for humans to make all the comparisons visually. Therefore, technological tools are essential for this task. In this work we use a pre-trained, publicly available Convolutional Neural Network (CNN) VGG19 that was trained on the ImageNet database. We adapted the VGG19 for the trademark image retrieval (TIR) task by fine tuning the network using two different databases. The VGG19v was trained with a database organized with trademark images using visual similarities, and the VGG19c was trained using trademarks organized by using conceptual similarities. The database for the VGG19v was built using trademarks downloaded from the WEB, and organized by visual similarity according to experts from the IP office. The database for the VGG19c was built using trademark images from the United States Patent and Trademarks Office and organized according to the Vienna conceptual protocol. The TIR was assessed using the normalized average rank for a test set from the METU database that has 922,926 trademark images. We computed the normalized average ranks for VGG19v, VGG19c, and for a combination of both networks. Our method achieved significantly better results on the METU database than those published previously.

The need to verify identity has become an everyday experience for most people. Biometrics is the principal means for reliable identification of people. Although iris recognition is the most reliable current technique for biometric identification, it has limitations because only segments of the iris are available due to occlusions from the eyelids, eyelashes, specular highlights, etc. The goal of this research is to study iris reconstruction from several 2D near infrared iris images, adding depth information to iris recognition. We expect that adding depth information from the iris surface will make it possible to identify people from a smaller segment of the iris. We designed a sensor for 2D near-infrared iris image acquisition. The method follows a pre-processing stage with the goal of performing iris enhancement, eliminating occlusions, reflections and extreme gray-level values, ending in iris texture equalization. The last step is the 3D iris model reconstruction based on several 2D iris images acquired at different angles. Results from each stage are presented.

Previous researchers have explored various approaches for predicting the gender of a person based on the features of the iris texture. This paper is the first to predict gender directly from the same binary iris code that could be used for recognition. We found that the information for gender prediction is distributed across the iris, rather than localized in particular concentric bands. We also found that using selected features representing a subset of the iris region achieves better accuracy than using features representing the whole iris region. We used the measures of mutual information to guide the selection of bits from the iris code to use as features in gender prediction. Using this approach, with a person-disjoint training and testing evaluation, we were able to achieve 89% correct gender prediction using the fusion of the best features of iris code from the left and right eyes.

There are many applications in image analysis where it is important to detect accurately patterns that include color and texture, e.g., plastic or concrete traffic barriers. This paper proposes a new method and extends a general machine vision approach for on-line pattern detection using color and textural information. Our proposed method includes the following steps: division of each image into sub-images, use of the Haralick and Binary Quaternion-Moment-Preserving methods to extract texture and color features, support vector machines for classification, and a post processing stage using clustering. The method was tested in two databases. The first one with three pattern types and the results yielded a detection rate of 96.4% with 14 false positives. The second database has nine pattern types and the results yielded a detection rate of 98.4% with 9 false positives. The results were compared advantageously with Haralick and BQMP methods separately.

Diabetic retinopathy is the leading cause of blindness in the working age population of the world. There is evidence that the digital image of the eye fundus is sensitive and specific for early signs of diabetic retinopathy. An automated image analysis method to detect early signs of diabetic retinopathy is highly desired for screening. The optic disc is one of the main structures visible in the eye fundus image that can be used as a reference to detect other abnormalities. In this paper, we propose a new cascade classifier based method for online optic disc detection. The cascade classifiers are trained using segmented images of optic discs and non-optic discs obtained from a training database (STARE). The method extracts Haar features from rectangular windows that are used to scan the digital image of the eye fundus. After training with segmented images from the STARE database, our optic disc detection method was tested with the DIARETDB1 database. Results showed high classification accuracy and fast detection capacity for online optic disc detection.

One important application of biometrics is determining the gender and age of customers so that attention could be specialized to improve sales. Gender classification has been a common topic of research and given the existence of face symmetry, determining gender based on half of the face seems feasible to significantly reduce computational cost. In this paper, we report the exploration of using the symmetrical characteristics of the face for representing and determining gender from only half of the face. We first divide the faces into two halves, and then select the best features separately from the left and right sides. The method uses 4 different mutual information measures to select features, minimum redundancy and maximal relevance (mRMR), normalized mutual information feature selection (NMIFS), conditional mutual information feature selection (CMIFS), and conditional mutual information maximization (CMIM). We tested our method on the FERET database using 5 fold cross-validation. It is shown that selection of features significantly improved gender classification accuracy compared to the use of full faces. We also show a significant reduction in processing time making real-time applications feasible.

In this paper, we report our extension of the use of feature selection based on mutual information and feature fusion to improve gender classification of face images. We compare the results of fusing three groups of features, three spatial scales, and four different mutual information measures to select features. We also showed improved results by fusion of LBP features with different radii and spatial scales, and the selection of features using mutual information. As measures of mutual information we use minimum redundancy and maximal relevance (mRMR), normalized mutual information feature selection (NMIFS), conditional mutual information feature selection (CMIFS), and conditional mutual information maximization (CMIM). We tested the results on four databases: FERET and UND, under controlled conditions, the LFW database under unconstrained scenarios, and AR for occlusions. It is shown that selection of features together with fusion of LBP features significantly improved gender classification accuracy compared to previously published results. We also show a significant reduction in processing time because of the feature selection, which makes real-time applications of gender classification feasible.

Face recognition has a wide range of possible applications in surveillance, access control, human computer interfaces and in electronic marketing and advertising for selected customers. Several models based on Gabor feature extraction have been proposed for face recognition with very good results on internationally available face databases. In this paper, we propose a methodological improvement to increase face recognition rate by selection and weighting Gabor jets by an entropy measure. We also propose improvements in the Borda count classification through a threshold to eliminate low score jets from the voting process to increase the face recognition rate. We show that combinations of weighting Gabor jets and threshold Borda yield the best results. We tested our methodological improvements on the FERET and the AR face databases. On the FERET database we reduce the total number of errors from 163 to 102 which is the highest score published up to date. The total number of errors in face recognition was reduced in 37%. On the AR database we also obtained important improvements and tested face images with illumination and gesticulation changes, and occlusions.

A filter method of feature selection based on mutual information, called normalized mutual information feature selection (NMIFS), is presented. NMIFS is an enhancement over Battiti's MIFS, MIFS-U, and mRMR methods. The average normalized mutual information is proposed as a measure of redundancy among features. NMIFS outperformed MIFS, MIFS-U, and mRMR on several artificial and benchmark data sets without requiring a user-defined parameter. In addition, NMIFS is combined with a genetic algorithm to form a hybrid filter/wrapper method called GAMIFS. This includes an initialization procedure and a mutation operator based on NMIFS to speed up the convergence of the genetic algorithm. GAMIFS overcomes the limitations of incremental search algorithms that are unable to find dependencies between groups of features.

Face detection methods are important in many applications such as face recognition, iris detection, eye-gaze estimation, gesture recognition and in man-machine interfaces, among others. In this paper, a new method for face detection of frontal faces is proposed. The method is based on a geometric transform we named Log-Linear Elliptic Transform, LET, that converts the lower semi-elliptical contour of the face into a vertical line in the transformed space. The vertical line can be detected by a simple projected histogram allowing a fast computation. The vertical line detection is sensitive to the face position, eccentricity and rotation. This allows the selection of an ellipse that better represents the face. The method was applied successfully in the detection of frontal faces in the FERET database.

Real-time face and iris detection on video sequences is important in diverse applications such as, study of the eye function, drowsiness detection, virtual keyboard interfaces, face recognition, and multimedia retrieval. In this paper, a real-time robust method is developed to detect irises on faces with coronal axis rotation within the normal range of to . The method allows head movements with no restrictions to the background. The method is based on anthropometric templates applied to detect the face and eyes. The templates use key features of the face such as the elliptical shape, and location of the eyebrows, nose, and lips. For iris detection, a template following the iris-sclera boundary shape is used. The method was compared to Maio-Maltoni's and Rowley's methods for face detection on five video sequences (TEST 1). The method was also assessed in an additional set of five video sequences for iris detection (TEST 2). Results of correct face detection in TEST 1 were above 99% in three of the five video sequences. The fourth video sequence reached 97.6% and the third 90.6%. In TEST 2, the iris detection was above 96% in all five video sequences with two above 99.7% and two at 100%. Face size estimation is also above 99.9%. The average processing time of our method was 0.02 s per frame. Thus, the proposed method can process frames at a rate near to 50 frames/s, and therefore, is applicable in real time in a standard personal computer (PC 1.8 GHz).

In this work we present a new method based on PSO (particle swarm optimization) to optimize templates for frontal face detection. In the past, several methods for face detection have been developed using face templates. These templates are based on common face features such as eyebrows, nose and mouth. Templates have been applied to a directional image containing faces computing a line integral to detect faces with high accuracy. In this paper, the PSO is used to select new templates optimizing its size and response to a face in the directional image. The method was tested on a database composed of two video sequences and compared to the results of the traditional anthropometric templates that contain features from the eyebrow, nose and mouth. Results show that templates selected by PSO have significant better performance in the estimation of face size and the line integral value. In both sequences face detection reached 99% and 100%. The templates have fewer number of points compared to the traditional anthropometric templates which will lead to lower processing time.

In this work, a new color image segmentation method, based on fuzzy min-max neural networks is presented. The proposed method is called FMMIS (fuzzy min-max neural network for image segmentation). The FMMIS method grows boxes from a set of seed pixels, to find the minimum bounded rectangle (MBR) for each object present in the images. The algorithm was tested on wood images of 10 defect categories and with images of frontal faces taken from the FERET database. The FMMIS algorithm outperformed alternative methods in terms of object detection rate, false positive detection rate, average execution time and the RUMA index. The proposed method is very fast and it may be applied to real-time image segmentation tasks.

This paper compares the classification performance of linear-system- and neural-network-based models in handwritten-digit classification and face recognition. In inputs to a linear classifier, nonlinear inputs are generated based on linear inputs, using different forms of generating products. Using a genetic algorithm, linear and nonlinear inputs to the linear classifier are selected to improve classification performance. Results show that an appropriate set of linear and nonlinear inputs to the linear classifier were selected, improving significantly its classification performance in both problems. It is also shown that the linear classifier reached a classification performance similar to or better than those obtained by nonlinear neural-network classifiers with linear inputs.

This paper proposes a new method for the design, through simulated evolution, of biologically inspired receptive fields in feedforward neural networks (NNs). The method is intended to enhance pattern recognition performance by creating new neural architectures specifically tuned for a particular pattern recognition problem. It proposes a combined neural architecture composed of two networks in cascade: a feature extraction network (FEN) followed by a neural classifier. The FEN is composed of several layers with receptive fields constructed by additive superposition of excitatory and inhibitory fields. A genetic algorithm (GA) is used to select receptive field parameters to improve classification performance. The parameters are receptive field size, orientation, and bias as well as the number of different receptive fields in each layer. Based on a random initial population where each individual represents a different neural architecture, the GA creates new enhanced individuals. The method is applied to handwritten digit classification and face recognition. In both problems, results show strong dependency between NN classification performance and receptive field architecture. GA selected parameters of the receptive fields produced improvements in the classification performance on the test set up to 90.8% for the problem of handwritten digit classification and up to 84.2% for the face recognition problem. On the same test sets, results were compared advantageously to standard feedforward multilayer perceptron (MLP) NNs where receptive fields are not explicitly defined. The MLP reached a maximum classification performance of 84.9% and 77.5% in both problems, respectively.

The purpose of the study is to compare neural network and linear system based models in 2-D pattern recognition tasks. Using a linear classifier, non-linear inputs are generated based on the linear inputs using different forms of generating products. These nonlinear inputs form a candidate set from which nonlinear inputs are selected to improve classification performance. A genetic search is performed to find an appropriate set of non-linear inputs. The method is applied to the handwritten digit recognition problem. Results show that the linear model with linear inputs reaches a classification performance on the testing database of 79.3%. However, when nonlinear inputs selected by the genetic algorithm were used, and included as new inputs, the classification performance increased up to 92.8%. These results are compared with those of three non-linear neural network models widely used in classification tasks using the same database. A single-layer perceptron with linear inputs reached 81.0% of correct classification. Perceptron models having a single-hidden layer and two-hidden layers reached classification results of 90.1% and 92.5%, respectively. Therefore, these results show that a linear classifier with an appropriate set of non-linear inputs reached a classification performance similar or better than those obtained by nonlinear neural network classifiers with one and two-hidden layer and with linear inputs.

A non-invasive interface to track eye position using digital image processing techniques is under development. Information about head and eye position is obtained from digital images. The objective is to develop an interface to detect eye position based only on digital image processing algorithms, free of electrodes or other electronic devices. We propose a method for eye tracking built into five stages. These include: coarse and fine face detection, finding the eye region of maximum probability, map of the pupil/iris location and pupil/iris detection. Using frontal face images obtained from a database, the probability maps for the eye region were built. Only gray levels are considered for this computation (8 bits). The algorithms for face and eye detection were assessed on 102 images from the Purdue database and on 897 images from a video sequence. The face detection algorithm reached a 99% and 100% correct detection rate on the databases respectively. On the same databases the pupil/iris detection algorithm reached 85.3% and 98.4% of correct detection respectively.

The paper describes a new method to incorporate biologically inspired receptive fields in feedforward neural networks to enhance pattern recognition performance. We propose a neural architecture composed of two networks in cascade: a feature extraction network followed by a neural classifier. A genetic algorithm is used to search for the receptive field configuration in the problem of handwritten digit recognition. The proposed network, with properly designed receptive fields, shows an improvement in the classification performance relative to other neural network models with fully connected architectures where receptive fields are not explicitly defined. A self organizing map is used to show the relative distance among patterns before and after the transformation performed by the network with biologically inspired receptive fields.

Applies modular neural network models to classify sleep-waking states in infants. The performances of three connectionist models are compared: (a) a multilayer perceptron (MLP), (b) a mixture of experts (ME) and (c) a fuzzy ganglionar lattice (FGL). We propose a new methodology for enhancing neural classifiers based on input variable selection and confusion error analysis using expert criteria. The ME model was more robust than the MLP and FGL models in the presence of inconsistent or noisy data. Input variable selection and confusion error analysis using expert criteria led to parsimonious models with less parameters and better classification rates.

The article study compares neural network based models and difference equation based models in 2D pattern recognition of handwritten digits. Pattern classifier models based on two neural network models were implemented using multi-layer feed forward neural networks: single layer perceptron and a multi-layer perceptron with one hidden layer. These two neural network models have nonlinear sigmoidal threshold units. A classifier model based on difference equations with linear weighted sum of all inputs was also implemented. Beside linear inputs, nonlinearities were introduced in all of the models by including products of inputs as new inputs. All models were trained/adjusted and tested using a handwritten digit database composed of a training set of 2360 digits and a testing set of 1320 digits. Each handwritten digit is composed of 15/spl times/23 binary pixels and each pixel is considered as a single input to the classifier. Classifier performance was measured as the correct classification rate on the testing database. Results show that the difference equation model with only linear inputs yielded the worse results (72.4%). Nevertheless, as products among inputs were included as new inputs, the classification performance increased (84.1%), thus exceeding the results of the single-layer perceptron with linear inputs (72.7%). The single-layer perceptron with nonlinear inputs gave an improvement of the classification result to 86.3%. The products among inputs can be interpreted as convolutional filtering over the input image to detect lines in different directions producing an overall improvement in the classification performance of all models.

A method is developed to classify handwritten numbers based on prototypes created using Euclidean distance, weighted voting among closest prototypes and fuzzy rules to solve confusions. The set of closest prototypes is determined by an acceptance distance. The fuzzy rules use specialized functions to measure characteristics on the handwritten digits to determine whether a digit belongs to a particular class. Classification performance is compared among following approaches: closest prototype, weighted voting including linear and exponential weighting, and voting plus fuzzy rules. The method is also compared to an algorithm based on a multilayer perceptron (MLP) network with augmented training. The best classification achieved was by voting plus fuzzy rules (96.3/spl plusmn/0.4% for 11 simulations). This result compares favorably with those obtained by MLP on the same testing database (94.6/spl plusmn/0.5%).

In this paper modular neural networks are used to improve handwritten digit recognition. To evaluate the performance of modular networks, a comparison is made with a global neural network, on the same database. Two basic kind of modular networks are considered: 1) seven expert modular networks in which five of them are provided for digits 0, 1, 2, 5, 6, 7 and the rest for the pair of digits 3-8 and 4-9 respectively; and 2) a modular neural network with an expert module for each feature extracted from the handwritten digit image. The cooperation is among modules extracting slope and radial projection from each digit. Two type of cooperation among modular networks are considered: neural network and weighted combination of the modules outputs. The models were trained and tested on a different set of digits. The results show that by using modular network for features, it is possible to improve classification performance on handwritten digits, from 91.0% in the case of global networks to 93.5% of modular networks.