Although stereo matching algorithms based on belief propagation (BP) tend to show excellent matching performance, their huge computational complexity has been the major barrier to real-time applications. In this light, we propose a parallel very large scale integration (VLSI) architecture for BP computation, which has only simple integer operations and shows low matching error rate for the Middlebury database.

We propose a multistep statistical procedure to determine the confidence interval of the number of features that should be retained in appearance-based face recognition, which is based on the eigen decomposition of covariance matrices. In practice, due to sampling variation, the empirical eigenpairs differ from their underlying population counterparts. The empirical distribution is difficult to derive, and it deviates from the asymptotic approximation when the sample size is limited, which hinders effective feature selection. Hence, we propose a new technique, MIZM (modified indifference zone method), to estimate the confidence interval of the number of features. MIZM overcomes the singularity problem in face recognition and extends the indifference zone selection from PCA to LDA. The simulation results on the ORL, UMIST, and FERET databases show that the overall recognition performance based on MIZM is improved from that using all available features or heuristically selected features. The relatively small number of features also indicates the efficiency of the proposed feature selection method. MIZM is motivated by feature selection for face recognition, but it extends the indifference zone method from PCA to LDA and can be applied in general LDA tasks.

We present a hybrid approach to face feature extraction based on the trace transform and the novel kernel partial leastsquares discriminant analysis (KPA). The hybrid approach, called trace kernel partial least-squares discriminant analysis (TKPA), first uses a set of 15 trace functionals to derive robust and discriminative facial features and then applies the KPA method to reduce their dimensionality. The feasibility of the proposed approach was successfully tested on the XM2VTS database, where a false rejection rate (FRR) of 1.25% and a false acceptance rate (FAR) of 2.11% were achieved in our best-performing face authentication experiment. The experimental results also show that the proposed approach can outperform kernel methods such as generalized discriminant analysis (GDA), kernel Fisher analysis (KFA), and complete kernel Fisher discriminant analysis (CKFA) as well as combinations of these methods with features extracted using the trace transform.

We describe a framework aimed at performing facebased biometric user authentication for Web resources through client–server secured sessions. A novel front end for face video sequences processing is developed in which face detection and shot selection are performed at the client side while statistical multishot pose-corrected face verification is performed at the server side. We explain all the image processing steps, from acquisition through decision, paying special attention to a PDM-based pose correction subsystem and a GMM-based sequence decision test. The pose correction relies on projecting a face-shape mesh onto the set of PDM eigenvectors and back-projecting it after changing the coefficients associated with pose variation. The aligned texture features compose the observation vectors ready to be plugged into a GMMbased likelihood ratio for statistical decision. The pose correction strategy has been previously validated over the XM2VTS and CMUPIE data sets, the GMM-based video sequence decision test has been compared to other video-based systems using the BANCA data set, and the complete proposed system has been tested on a new video data set from the European Network of Excellence BIOSECURE.

We present a new system of near-infrared capture that aims to eliminate ambient light effects. We design this sensor to be integrated into the mobile platform Vérification d’identité Numérique Sécurisé Itinérante (VINSI). One of the originalities of this sensor is that we can acquire two images at the same time: one in infrared and the other in visible light. We develop a novel method for landmark detection of sufficiently low complexity, so that it can be implemented on our mobile platform. We test this method on three different near-infrared face databases, and we observe stability in the precision of localization over these databases associated to performance comparable to the state of the art.

The accuracy of a three-dimensional (3-D) face recognition system depends on a correct registration that aligns the facial surfaces and makes a comparison possible. The best results obtained so far use a costly one-to-all registration approach, which requires the registration of each facial surface to all faces in the gallery. We explore the approach of registering the new facial surface to an average face model (AFM), which automatically establishes correspondence to the preregistered gallery faces. We propose a new algorithm for constructing an AFM and show that it works better than a recent approach. We inspect thin-plate spline and iterative closest-point-based registration schemes under manual or automatic landmark detection prior to registration. Extending the single-AFM approach, we consider employing categoryspecific alternative AFMs for registration and evaluate the effect on subsequent classification. We perform simulations with multiple AFMs that correspond to different clusters in the face shape space and compare these with gender- and morphology-based groupings. We show that the automatic clustering approach separates the faces into gender and morphology groups, consistent with the other race effect reported in the psychology literature. Last, we describe and analyze a regular resampling method, that significantly increases the accuracy of registration.

Traditionally, the biometric field has viewed the subject as a passive source of the biometric sample rather than as an interactive component of the biometric system. But fingerprint image quality is highly dependent on the human–computer interaction and usability considerations of the acquisition system. Those factors impacting the acquisition of high-quality images must be identified, and real-time feedback for subjects to ensure acceptable quality images must be integrated into fingerprint capture systems. We report on a usability study that examined the influence of instructional materials on the user (subject) performance of a 10-print slap acquisition process. In addition, we also investigated the relationship of pressure and image quality as a mechanism to provide real-time feedback to the subject. The usability study included 300 participants who received instructions and interacted with the scanner. How information is provided to the subject on interacting with the fingerprint device does indeed affect image quality. The pressure findings are less conclusive; there was no clear relationship between image quality and pressure that could be exploited for feedback to the subject. However, a minimum pressure was required to initiate our capture process.

Although many image quality measures have been proposed for fingerprints, few works have taken into account how differences among capture devices impact the image quality. Several representative measures for assessing the quality of fingerprint images are compared using an optical and a capacitive sensor. We implement and test a representative set of measures that rely on different fingerprint image features for quality assessment. The capability to discriminate between images of different quality and the relationship with the verification performance are studied. For our verification experiments, we use minutiae- and ridge-based matchers, which are the most common approaches for fingerprint recognition. We report differences depending on the sensor, and interesting relationships between sensor technology and features used for quality assessment are also pointed out.

Recent research has shown that it is possible to spoof a variety of fingerprint scanners using some simple techniques with molds made from plastic, clay, Play-Doh, silicon, or gelatin materials. To protect against spoofing, methods of liveness detection measure physiological signs of life from fingerprints, ensuring that only live fingers are captured for enrollment or authentication. We propose a new liveness detection method based on noise analysis along the valleys in the ridge-valley structure of fingerprint images. Unlike live fingers, which have a clear ridge-valley structure, artificial fingers have a distinct noise distribution due to the material’s properties when placed on a fingerprint scanner. Statistical features are extracted in multiresolution scales using the wavelet decomposition technique. Based on these features, liveness separation (live/ nonlive) is performed using classification trees and neural networks. We test this method on the data set, that contains about 58 live, 80 spoof (50 made from Play-Doh and 30 made from gelatin), and 25 cadaver subjects for 3 different scanners. We also test this method on a second data set that contains 28 live and 28 spoof (made from silicon) subjects. Results show that we can get approximately 90.9– 100% classification of spoof and live fingerprints. The proposed liveness detection method is purely software-based, and application of this method can provide antispoofing protection for fingerprint scanners.

We propose a phase-encoding-based digital watermarking technique for fingerprint template protection and verification. We extract a signature from the one-dimensional Fourier phase of the original fingerprint image (template) and then embed it back into the image using a variation of phase-shift keying modulation and the spread-spectrum method. To minimize the degradation to important minutiae features of fingerprint data, we segment the template into a region of interest (ROI) and a region of background (ROB) and embed the watermark adaptively in these two regions using an adaptive phase quantization method. We give experimental results with fingerprint biometric data that demonstrate the protection of fingerprint templates as well as the verification of fingerprint recognition.

Biometrics is rapidly becoming the principal technology for automatic people authentication. The main advantage in using biometrics over traditional recognition approaches relies in the difficulty of losing, stealing, or copying individual behavioral or physical traits. The major weakness of biometrics-based systems relies in their security: in order to avoid data stealing or corruption, storing raw biometric data is not advised. The same problem occurs when biometric templates are employed, since they can be used to recover the original biometric data. We employ cryptographic techniques to protect dynamic signature features, making it impossible to derive the original biometrics from the stored templates, while maintaining good recognition performances. Together with protection, we also guarantee template cancellability and renewability. Moreover, the proposed authentication scheme is tailored to the signature variability of each user, thus obtaining a user adaptive system with enhanced performances with respect to a nonadaptive one. Experimental results show the effectiveness of our approach when compared to both traditional nonsecure classifiers and other, already proposed protection schemes.

We present a pioneering study on machine dating of manuscripts written by an individual. Analysis of handwriting style forms the core of the proposed method. A general framework is presented for automatic time stamping of handwritten manuscripts. Initially, it is hypothesized that a manuscript can be dated, to a certain level of accuracy, by looking at the way it is written. The hypothesis is then verified with real samples of known dates. Experiments on a database containing manuscripts of Gustave Flaubert (1821– 1880), the famous French novelist, reports about 62% accuracy while dating the manuscripts within a range of five calendar years with respect to their exact year of writing.

We propose a new user authentication system based on spoken signatures, where online signature and speech signals are acquired simultaneously. The main benefit of this multimodal approach is better accuracy at no extra cost for the user in terms of access time or inconvenience. Another benefit lies in a better robustness against intentional forgeries due to the extra difficulty for the forger to produce both signals. We set up an experimental framework to measure these benefits on MyIDea, a realistic multimodal biometric database publicly available. More specifically, we evaluate the performance of state of the art modeling systems based on Gaussian mixture models (GMM) and hidden Markov models (HMM) applied independently to the pen and voice signal, where a simple rule-based score fusion procedure is used. We conclude that the best performance is achieved by the HMMs, provided that their topology is optimized on a per user basis. Furthermore, we show that more precise models can be obtained through the use of maximum a posteriori probability (MAP) training instead of the classically used expectation maximization (EM). We also measure the impact of multisession scenarios versus monosession scenarios, and the impact of skilled versus unskilled signature forgeries attacks.

We consider the problem of authentication of biometric identification documents via mobile devices such as mobile phones or personal digital assistants (PDAs). We assume that the biometric identification document holds biometric data (e.g., face or fingerprint) in the form of an image and personal data in the form of text, both being printed directly onto the identification document. The proposed solution makes use of digital data hiding in order to crossstore the biometric data inside the personal data and vice versa. Moreover, a theoretical framework is presented that should enable analysis and guide the design of future authentication systems based on this approach. In particular, we advocate the separation approach, which uses robust visual hashing techniques in order to match the information rates of biometric and personal data to the rates offered by current image and text data hiding technologies. We also describe practical schemes for robust visual hashing and digital data hiding that can be used as building blocks for the proposed authentication system. The obtained experimental results show that the proposed system constitutes a viable and practical solution.

The human iris is perhaps the most accurate biometric for use in identification. Commercial iris recognition systems currently can be found in several types of settings where a person’s true identity is required: to allow passengers in some airports to be rapidly processed through security; for access to secure areas; and for secure access to computer networks. The growing employment of iris recognition systems and the associated research to develop new algorithms will require large databases of iris images. If the required storage space is not adequate for these databases, image compression is an alternative. Compression allows a reduction in the storage space needed to store these iris images. This may, however, come at a cost: some amount of information may be lost in the process. We investigate the effects of image compression on the performance of an iris recognition system. Compression is performed using JPEG-2000 and JPEG, and the iris recognition algorithm used is an implementation of the Daugman algorithm. The imagery used includes both the CASIA iris database as well as the iris database collected by the University of Bath. Results demonstrate that compression up to 50:1 can be used with minimal effects on recognition.

We investigate the potential of foot biometric features based on geometry, shape, and texture and present algorithms for a prototype rotation invariant verification system. An introduction to origins and fields of application for footprint-based personal recognition is accompanied by a comparison with traditional hand biometry systems. Image enhancement and feature extraction steps emphasizing specific characteristics of foot geometry and their permanence and distinctiveness properties, respectively, are discussed. Collectability and universality issues are considered as well. A visualization of various test results comparing discriminative power of foot shape and texture is given. The impact on real-world scenarios is pointed out, and a summary of results is presented.

In the biometric verification system of a smart gun, the rightful user of the gun is recognized based on grip-pattern recognition. It was found that the verification performance of grip-pattern recognition degrades strongly when the data for training and testing the classifier, respectively, have been recorded in different sessions. The major factors that affect the verification performance of this system are the variations of pressure distribution and hand position between the probe image and the gallery image of a subject. In this work, three methods are proposed to reduce the effect of the variations by using different sessions for training, image registration, and classifier fusion. Based on these methods, the verification results are significantly improved.

We provide a survey of hand biometric techniques in the literature and incorporate several novel results of hand-based personal identification and verification. We compare several feature sets in the shape-only and shape-plus-texture categories, emphasizing the relevance of a proper hand normalization scheme in the success of any biometric scheme. The preference of the left and right hands or of ambidextrous access control is explored. Since the business case of a biometric device partly hinges on the longevity of its features and the generalization ability of its database, we have tested our scheme with time-lapse data as well as with subjects that were unseen during the training stage. Our experiments were conducted on a hand database that is an order of magnitude larger than any existing one in the literature.

The target recognition ability of the human visual system can be severely affected by vibration of the imaging system. Although vibration can be minimized by proper design, it is often the main limitation on visual perception despite the best attempts at stabilization. There are two types of image degradation caused by vibration of the camera. The first is the vibration of the line of sight (LOS), causing location changes of the scene in successive frames. The second degradation type is the blur induced in each frame of the sequence due to motion during the exposure. We investigate the relative effect of these two degradation sources on the human visual system recognition abilities by conducting a series of psychophysical experiments. These experiments show that object recognition and orientation recognition abilities may be affected more by the motion blur of each frame than by the oscillation of the scene. An important implication of these results on digital image sequence restoration algorithms is that under certain conditions, the main effort should be put toward motion deblurring rather than on the precise registration of the frames.

Electronic visual prostheses, or “bionic eyes,” are likely to provide some coarse visual sensations to blind patients who have these systems implanted. The quality of artificially induced vision is anticipated to be very poor initially. Research described explores image processing techniques that improve perception for users of visual prostheses. We describe visual perception experiments undertaken with 192 normally sighted viewers to simulate artificially induced vision expected from emerging electronic visual prosthesis designs. Several variations of region-of-interest (ROI) processing were applied to images that were presented to subjects as low-resolution 25×25 binary images. Several additional processing methods were compared to determine their suitability for use in automatically controlling a zoom-type function for visual prostheses. The experiments show that ROI processing improves scene understanding for low-quality images when used in a zoom application.

Color printer calibration is the process of deriving correction functions for device signals (e.g., CMYK), so that the device can be maintained with a fixed known characteristic color response. Since the colorimetric response of the printer can be a strong function of the halftone, the calibration process must be repeated for every halftone supported by the printer. The effort involved in the calibration process thus increases linearly with the number of halftoning methods. In the past few years, it has become common for high-end digital color printers to be equipped with a large number of halftones, thus making the calibration process onerous. We propose a halftone independent method for correcting color (CMY or CMYK) printer drift. Our corrections are derived by measuring a small number of halftone independent fundamental binary patterns based on the 22 binary printer model by Wang et al. Hence, the required measurements do not increase as more halftoning methods are added. First, we derive a halftone correction factor (HCF) that exploits the knowledge of the relationship between the true printer response and the 22-model predicted response for a given halftoning scheme. Therefore, the true color drift can be accurately predicted from halftone-independent measurements and corrected correspondingly. Further, we develop extensions of our proposed color correction framework to the case when the measurements of our fundamental binary patches are acquired by a common desktop scanner. Finally, we exploit the application of the HCF to correct color drift across different media (papers) and for halftoneindependent spatial nonuniformity correction.

Human visual system (HVS) modeling has become a critical component in the design of digital halftoning algorithms. Methods that exploit the characteristics of the HVS include the direct binary search (DBS) and optimized tone-dependent halftoning approaches. The spatial sensitivity of the HVS is low-pass in nature, reflecting the physiological characteristics of the eye. Several HVS models have been proposed in the literature, among them, the broadly used Näsänen’s exponential model, which was later shown to be constrained in shape. Richer models are needed to attain better halftone attributes and to control the appearance of undesired patterns. As an alternative, models based on the mixture of bivariate Gaussian density functions have been proposed. The mathematical characteristics of the HVS model thus play a key role in the synthesis of model-based halftoning. In this work, alpha stable functions, an elegant class of functions richer than mixed Gaussians, are exploited to design HVS models to be used in two different contexts: monochrome halftoning over rectangular and hexagonal sampling grids. In the two scenarios, alpha stable models prove to be more efficient than Gaussian mixtures, as they use less parameters to characterize the tails and bandwidth of the model. It is shown that a decrease in the model’s bandwidth leads to homogeneous halftone patterns, and conversely, models with heavier tails yield smoother textures. These characteristics, added to their simplicity, make alpha stable models a powerful tool for HVS characterization.

We propose a new finite mixture model based on the formalism of general Gaussian distribution (GGD). Because it has the flexibility to adapt to the shape of the data better than the Gaussian, the GGD is less prone to overfitting the number of mixture classes when dealing with noisy data. In the first part of this work, we propose a derivation of the maximum likelihood estimation for the parameters of the new mixture model, and elaborate an information-theoretic approach for the selection of the number of classes. In the second part, we validate the proposed model by comparing it to the Gaussian mixture in applications related to image and video foreground segmentation.

H.264/AVC employs variable-size motion compensation with multiple reference frames that can significantly improve the coding performance compared with the previous video coding standards. However, the computational complexity of the H.264/AVC encoder increases dramatically due to the various size modes used. One of the most challenging problems in implementing this encoder is mode decision. We propose a novel intermode decision algorithm (NIMDA) to reduce the candidate mode set, which utilizes two basic elements, the sum absolute difference (SAD) of each 44 block from the initial test of the inter-1616 and the texture characteristic. Simulation results show that our proposed algorithm can save about 46% in computational complexity, with negligible loss of coding efficiency.

One of the main image representations in mathematical morphology is the shape decomposition representation, useful for image compression and pattern recognition. The morphological shape decomposition representation can be generalized to extend the scope of its algebraic characteristics as much as possible. With these generalizations, the morphological shape decomposition (MSD) role to serve as an efficient image decomposition tool is extended to interpolation of images. We address the binary and grayscale interframe interpolation by means of generalized morphological shape decomposition. Computer simulations illustrate the results.

We propose a solution for the computer-aided reconstruction of strip-cut shredded documents. First of all, the visual content of the strips is automatically extracted and represented by a number of numerical features. Usually, the pieces of different pages have been mixed. A grouping of the strips belonging to a same page is thus realized by means of a clustering operator, to ease the successive matching performed by a human operator with the help of a computer.

We integrate stylized rendering with an efficient multiresolution image representation, enabling a user to control how compression affects the aesthetic appearance of an image. We adopt a point-based rendering approach to progressive image transmission and compression. We use a novel, adaptive farthest point sampling algorithm to represent the image at progressive levels of detail, balancing global coverage with local precision. A progressively generated discrete Voronoi diagram forms the common foundation for our sampling and rendering framework. This framework allows us to extend traditional photorealistic methods of image reconstruction by scattered data interpolation to encompass nonphotorealistic rendering. It supports a wide variety of artistic rendering styles based on geometric subdivision or parametric procedural textures. Genetic programming enables the user to create original rendering styles through interactive evolution by aesthetic selection. We compare our results with conventional compression, and we discuss the implications of using nonphotorealistic representations for highly compressed imagery.

Dust, scratches, or hair on originals (prints, slides, or negatives) distinctly appear as light or dark artifacts on a scan. These unsightly artifacts have become a major consumer concern. There are several scenarios for removal of dust and scratch artifacts. One scenario is during acquisition, e.g., while scanning photographic media. Another is artifact removal from a digital image in an image editor. For each scenario, a different solution is suitable, with different performance requirements and differing levels of user interaction. This work describes a comprehensive set of algorithms for automatically removing dust and scratches from images. Our algorithms solve a wide range of use scenarios. A dust and scratch removal solution has two steps: a detection step and a reconstruction step. Very good detection of dust and scratches is possible using side information, such as provided by dedicated hardware. Without hardware assistance, dust and scratch removal algorithms generally resort to blurring, thereby losing image detail. We present algorithmic alternatives for dust and scratch detection. In addition, we present reconstruction algorithms that preserve image detail better than previously available alternatives. These algorithms consistently produce visually pleasing images in extensive testing.

Mirroring an image (horizontally or vertically) and rotation by 90, 180, and 270 deg in the spatial domain is implemented by changing the signs of the 2-D discrete cosine transform (DCT) coefficients of the original image appropriately. This approach leads to an efficient compressed domain-based image mirroring and rotation. This method is now extended to a 2-D discrete sine transform (DST), i.e., a 2-D DST possesses properties similar to 2-D DCT and it can be applied to image mirroring (horizontally or vertically) and also rotation by angles of 90, 180, and 270 deg. We illustrate these methods in a video-editing application.

The blocking effect is the major drawback in discrete cosine transform (DCT)-based codecs at low bit rates. When a high compression ratio is required, boundaries between adjacent image blocks become visible. This kind of degradation probably will affect the judgment of an end user. In this work, a four-neighboring-block zero-masking technique is proposed in the DCT frequency domain. The proposed algorithm uses a shift block within four adjacent DCT blocks to reduce computational complexity. By slightly modifying several DCT coefficients in the shift block, the artifacts that result from quantization and dequantization processes are alleviated. The simulation results show that both visual perception and objective image quality are noticeably improved.

A new spatially adaptive shrinkage approach based on the nonsubsampled contourlet transform (NSCT) to despeckling synthetic aperture radar (SAR) images is proposed. This method starts from the existing stationary wavelet transform (SWT)–domain Gamma-exponential likelihood model combined with a local spatial prior model and extends the model further for despeckling an SAR image via spatially adaptive shrinkage in the NCST domain. The proposed NSCT-domain shrinkage estimator consists of a new likelihood ratio function and a new prior ratio function, both of which are dependent on the estimated masks for the NSCT coefficients. The former is established by the Gamma distribution with variable scale and shape parameters and the exponential distribution with variable scale parameter to adapt the shrinkage estimator to the redundancy property of the NSCT. Parameters of these two distributions are estimated by using moment-based estimators. The latter is equipped with directional neighborhood configurations to accommodate the estimator to the flexible directionality of the NSCT, and thus to enhance the detail fidelity. We validate the proposed method on real SAR images and demonstrate the excellent despeckling performance through comparisons with the SWT-based counterpart, two classical spatial filters, and the contourlet transform-based despeckling technique.

Many magnification algorithms have been proposed in past decades, most of which concentrate on the smooth reconstruction of edge structures. Edge reconstruction, however, can destroy corners, thus producing perceptually unpleasant rounded corner structures. In this work, corner shock filtering is designated for enhancing corners relative to the known edge shock filtering, based on a new measure of corner strength and the theory of level-sets motion under curvature. By combining directional diffusion, edge shock filtering, and corner shock filtering, a regularized partial differential equation (PDE) approach for magnification is proposed to simultaneously reconstruct the edges and preserve the corners. Moreover, the proposed PDE approach is also robust to random noises. Experimental results in both cases of grayscale and color images confirm the effectiveness of our approach.

We propose a novel B-spline active contour model based on image fusion. Compared with conventional active contours, this active contour has two advantages. First, it is represented by a cubic B-spline curve, which can adaptively determine the curve parameter’s step length; and it can also effectively detect and express the object contour’s corner points. Second, it is implemented in connection with image fusion. Its external image force is modified as the weighted sum of two modal image forces, with the two weights in terms of a local region’s image entropy or image contrast’s standard deviation. The experiments indicate that this active contour can accurately detect both the object’s contour edge and the corner points. Our experiments also indicate that the active contour’s convergence with a weighted image force by the image contrast’s standard deviation is more accurate than that of image entropy, restraining the influence of the texture or pattern.

A new method is presented for automatic contour extraction in computerized tomography (CT) reconstruction, based on projection segmentation and backprojection. Contour extraction is of great importance in medical applications, especially in nondestructive inspection. Manual segmentation includes pixel selection, geometrical boundary selection, and tracing, however, it depends on the experience of the operators and is time-consuming. Most fully automatic segmentation methods for CT images are usually impractical due to the obvious noise in reconstructed slices. Because the backprojection algorithm is of a high-pass nature, the noise is always in high frequencies. The projections, however, are much clearer than the reconstructed slices. The advantage of the new algorithm is its simplicity in segmenting the projections instead of the slices. Simulation results illustrate the accuracy of this method. A circular trajectory cone-beam micro-CT system is used to validate the algorithm. The experiment result shows a perfect 3-D contour of a rat skull.

This work proposes a new approach using a committee machine of artificial neural networks to classify masses found in mammograms as benign or malignant. Three shape factors, three edge-sharpness measures, and 14 texture measures are used for the classification of 20 regions of interest (ROIs) related to malignant tumors and 37 ROIs related to benign masses. A group of multilayer perceptrons (MLPs) is employed as a committee machine of neural network classifiers. The classification results are reached by combining the responses of the individual classifiers. Experiments involving changes in the learning algorithm of the committee machine are conducted. The classification accuracy is evaluated using the area Az under the receiver operating characteristics (ROC) curve. The Az result for the committee machine is compared with the Az results obtained using MLPs and single-layer perceptrons (SLPs), as well as a linear discriminant analysis (LDA) classifier. Tests are carried out using the student’s t-distribution. The committee machine classifier outperforms the MLP, SLP, and LDA classifiers in the following cases: with the shape measure of spiculation index, the Az values of the four methods are, in order, 0.93, 0.84, 0.75, and 0.76; and with the edge-sharpness measure of acutance, the values are 0.79, 0.70, 0.69, and 0.74. Although the features with which improvement is obtained with the committee machines are not the same as those that provided the maximal value of Az (Az=0.99 with some shape features, with or without the committee machine), they correspond to features that are not critically dependent on the accuracy of the boundaries of the masses, which is an important result.

We present an automatic and real-time vision system to inspect the two sides of a smart card module using fuzzy logic. In the smart card industry, the module is the referential element supporting the chip. It is designed via successive microelectronic operations to be finally embedded in the plastic card. For the chip side, the system can detect overflow of resin, integrity problems of the surface or of the resin protection, and misalignments. For the contact side, the system can detect all the defects affecting the module surface such as scratches, marks, or pollution. The vision system is flexible and real time. It is organized in dedicated layers and offers the necessary elements including controlled illumination systems to manage different films. A priori knowledge is introduced in the system, and fuzzy classifiers allow fast and reliable decisions. This industrial system has exhibited an efficiency comparable to that of experienced human observers but at much higher speed.

Viewing angle performance in in-plane switching liquid crystal displays (IPS-LCDs) has been greatly improved, using two technologies. One is optical compensation technology using a biaxial film. The other is a newly developed IPS-Pro cell structure with higher transmission efficiency. These technologies are successfully introduced into the fabrication of 32-in. IPS-LCDs with a minimum contrast ratio over three times that of conventional IPS-LCDs and with color saturation (area ratio to NTSC at CIE1931 xy chromaticity coordinates) of more than 70% at almost all viewing angles.