This PDF file contains the editorial “Quality Control by Artificial Vision” for JEI Vol. 13 Issue 03

Superquadrics are able to represent a large variety of objects with only a few parameters and a single equation. We present a superquadric representation strategy for automotive parts composed of 3-D triangle meshes. Our strategy consists of two major steps of part decomposition and superquadric fitting. The originalities of this approach include the following two features. First, our approach can represent multipart objects with superquadrics successfully by applying part decomposition. Second, superquadrics recovered from our approach have the highest confidence and accuracy due to the 3-D watertight surfaces utilized. A novel, generic 3-D part decomposition algorithm based on curvature analysis is also proposed. Experimental results demonstrate that the proposed part decomposition algorithm is able to segment multipart objects into meaningful single parts efficiently. The proposed superquadric representation strategy can then represent each individual part of the original objects with a superquadric model successfully.

A fabric's tendency to wrinkle is vitally important to the textile industry as it impacts the visual appeal of apparel. Current methods of grading this characteristic, called fabric smoothness, are very subjective and inadequate. As such, a quantitative method for assessing fabric smoothness is of the utmost importance to the textile community. To that end, we propose a laser-based surface-profiling system that utilizes a smart camera to sense the 3-D topography of fabric specimens. The system incorporates methods based on anisotropic diffusion and the facet model for characterizing edge information that ultimately relate to a specimen's degree of wrinkling. We detail the initial steps in a large-scale validation of this system. Using histograms of the extracted features, we compare the output of the system between two studies that total more than 200 fabric specimens. The results show that with the features used so far, this system is at least as good as the current American Association of Textile Chemists and Colorists (AATCC) smoothness grading system.

We extend our previous work on the image segmentation of electronic structures on patterned wafers to improve the defect detection process on optical inspection tools. Die-to-die wafer inspection is based on the comparison of the same area on two neighboring dies. The dissimilarities between the images are a result of defects in this area of one of the dies. The noise level can vary from one structure to the other, within the same image. Therefore, segmentation is required to create a mask and apply an optimal threshold in each region. Contrast variation on the texture can affect the response of the parameters used for the segmentation. We show a method to anticipate these variations with a limited number of training samples, and modify the classifier accordingly to improve the segmentation results.

Dual-energy techniques are used to carry out selective and quantitative imaging of materials consisting of two components. We present a technique that takes advantage of a deterministic simulation tool to find the optimal x-ray spectra used in a dual-energy system and to calibrate the measurement protocol. To optimize the choice of energy spectra, a signal-to-noise ratio (SNR) criterion on the estimated thickness of the materials is established. To further study the reliability of the SNR criterion, it is first compared to the measurement quality, expressed by a contrast-to-noise ratio, which characterizes the input images. Then, the numerical conditioning of the linear dual-energy system of equations is investigated to probe the stability of the inversion. Once the choice of energy spectra is settled, the reconstructed values of the thickness are modeled as third-order polynomials expressed in terms of dual-energy measurements. An application to glass-wool materials is presented.

The inspection and monitoring of the wear of grinding tools is essential to ensure the quality of the grinding tool surface and the finished product. Most of the current methods for examining a grinding tool surface rely on dismounting the grinding tool. Often, the state of the grinding tool surface is checked indirectly by evaluating the quality of the workpiece. We describe the application of image processing, which offers an effective means for in situ inspection and monitoring. It yields more detailed information about the surface and the kind of wear observed than the common methods. By using multidirectional illumination and image fusion, an image with a high degree of relevant information is generated that is then segmented using the wavelet transform (multiscale analysis) and classified to distinguish grains and cavities on the surface. Results of the application of the algorithms for a high-performance grinding tool with CBN grains embedded in a resin base are presented.

Industrial applications of computer vision, such as dimensional inspection, require, among other things, making automated procedures available for the analysis of gray-level image to derive application-dependent 3-D descriptions of their contents. In view of the quantitative 3-D reconstruction of an imaged object and to fully automate the evaluation process, we describe two segmentation and interpretation methods for the automated delineation of regions of interest belonging to the object and associated with free-form surfaces. The selection of the procedure to apply, either a mean shift based or level set based analysis, depends essentially on available a priori information relative to the localization and the shape of the object in the scene. The two approaches are part of a 3-D vision-based on-line inspection system. Results on images of manufactured parts acquired under realistic conditions illustrate the use of these two approaches.

For process control, linewidth measurements are commonly performed on semiconductor wafers using top-down images from critical dimension measurement scanning electron microscopes (CD-SEMs). However, a measure of the line sidewall shape will be required as linewidths continue to shrink. Sidewall shape can be measured by physically cleaving the device and performing an SEM scan of the cross section, but this process is time consuming and results in destruction of product wafers. We develop a technique to estimate sidewall shape from top-down SEM images using pattern recognition based on historical cross section/top-down image pairs. Features are computed on subimages extracted from the top-down images. Several combinations of principal component analysis (PCA) and flavors of linear discriminant analysis (LDA) are employed to reduce the dimensionality of the feature vectors and maximize the spread between different sidewall shapes. Direct, weighted LDA (DW-LDA) results in a feature set that provides the best sidewall shape estimation. Experimental testing of the sidewall estimation system shows a root mean square error of approximately 1.8% of the linewidth, showing that this system is a viable method for estimating sidewall shape with little impact on the fabrication process (no new hardware and a minimal increase in process setup).

The key to successful morphological image processing is the selection of structuring elements. There are a myriad of algorithms for a multitude of imaging applications, but in each and every instance, algorithm performance depends on the structuring elements. The classical approach to morphological processing is to have a human being, or a group of human beings, use intuition and an understanding of the goals to design algorithms based on erosions, openings, hit-or-miss transforms, and other basic morphological operators. Some of the methods are reviewed that have been developed for automatic algorithm design, where morphological operators are designed based on sample data, structural decomposition, and criteria set by the imaging scientist.

A new method based on quadratic constrained least-mean-square fitting to simultaneously determine both the best hyperbolic and elliptical fits to a set of scattered data is presented. Thus a linear solution to the problem of hyperbola-specific fitting is revealed for the first time. Pilu's method to fit an ellipse (with respect to distance) to observed data points is extended to select, without prejudice, both ellipses and hyperbolæ as well as their degenerate forms as indicated by optimality with respect to the algebraic distance. This novel method is numerically efficient and is suitable for fitting to dense datasets with low noise. Furthermore, it is deemed highly suited to initialize a better but more computationally costly least-square minimization of orthogonal distance. Moreover, Grassmannian coordinates of the hyperbolæ are introduced, and it is shown how these apply to fitting a prototypical hyperbola. Two new theorems on fitting hyperbolæ are presented together with rigorous proofs. A new method to determine the spatial uncertainty of the fit from the eigen or singular values is derived and used as an indicator for the quality of fit. All proposed methods are verified using numerical simulation, and working MATLAB^® programs for the implementation are made available. Further, an application of the methods to automatic industrial inspection is presented.

(TCP/IP) based, anonymous communication environment is introduced. The main system capabilities are central configurability and maintainability of the image processing nodes and the scalability of the framework. Furthermore a watchdog concept and IEC61131 conform signal integration comprising the Modbus/TCP protocol are integrated within the framework. A communication structure based on the tuplespace concept is presented. The possibility of integrating heterogeneous nodes is achieved by a consistent Java^TM implementation. As a proof of concept, a quality observation task at distributed observation points at a wire rolling mill is introduced. An image-processing algorithm based on least-squares fitting of parallel lines to observe quality relevant parameters is introduced and integrated as an exemplary machine vision implementation within the framework.

The study of rough textured surfaces such as road coverings, is generally made on gray-level images. This supposes that the variations of gray levels are representative of the local variations of the relief. This assumption is justified, in the case of surfaces that are uniformly colored, but finds its limit in the case when these surfaces present variations of color or aspect. The corresponding image then presents variations of gray levels that can be related to the color variations or to the relief variations or both. It becomes difficult in this case to work out the criteria of roughness based on image analysis. It is then necessary to develop, before any study of roughness, an estimation of the luminance map linked to color variations. To do that, we have linked the gray-level value to the height variation of the road-covering surface. Thus, we express the luminance distribution according to these height variations, obtained through a laser sensor, and to the image gray level. This method enables us to compute the distribution of the luminance map. We characterize this distribution by considering statistical parameters of its histogram. Then we test the effectiveness of our approach by comparing the evolution of the criterion of roughness on road surfaces, first without considering the luminance distribution and then by taking it into account. The results obtained show that this approach leads to a good discrimination by the criterion of roughness in the case of colored surfaces.

Very few image processing applications have dealt with x-ray luggage scenes in the past. Concealed threats in general, and low-density items in particular, pose a major challenge to airport screeners. A simple enhancement method for data decluttering is introduced. Initially, the method is applied using manually selected thresholds to progressively generate decluttered slices. Further automation of the algorithm, using a novel metric based on the Radon transform, is conducted to determine the optimum number and values of thresholds and to generate a single optimum slice for screener interpretation. A comparison of the newly developed metric to other known metrics demonstrates the merits of the new approach. On-site quantitative and qualitative evaluations of the various decluttered images by airport screeners further establishes that the single slice from the image hashing algorithm outperforms traditional enhancement techniques with a noted increase of 58% in low-density threat detection rates.

We describe an artificial vision system for the quality control of cherries. We develop image-processing software determining three types of information describing the quality of the fruit: the color as an indicator of ripeness, the presence of defects such as cracking, and the size. The sorting of cherries conditioned by all these criteria is carried out at a high cadence (20 cherries/s). Real-time image processing is then necessary. A simple sensor is used to synchronize the processing and the sorting of cherries. Air actuators are used to eject them after quality control by vision. We present the architecture of the developed system. We show its efficiency through experimental results and we give some perspectives of improvement.

Multispectral imagery is a large domain with a number of practical applications: thermography and quality control in industry, food science, and agronomy. The main interest is to obtain spectral information of the objects for which a reflectance signal can be associated to physical, chemical, and/or biological properties. Agronomic applications of multispectral imagery generally involve the acquisition of several images in visible and near infrared wavelengths. We first present a different kind of multispectral devices used for agronomic applications and then introduce an original multispectral acquisition system based on a single CCD. First results in laboratory are detailed, presenting a detection method using a neural network and in-field acquisitions and their results are shown. To improve the quality of weed detection, the spatial distribution of crops is used by a second method. Finally, the first works on merging are outlined.

Animals accomplish color vision by characterizing the scene in two or more spectrally overlapping bands, normalizing to minimize brightness effects, computing spectral discriminants (called hues) from the normalized readings, and attributing those hues to the scene with the brightness variations restored in the "colored image." We explore doing the same process in forming images for camera systems. To make experimental testing simpler, we use available spectrally overlapping band cameras—"color" CCD cameras—whose bands show considerable overlap but were not chosen for the task to which we employ them. We use these bands and a very small number of samples from each class of object to form a single artificial color—"green pepper but not snow peas and not carrots." An image of a plate of such vegetables in that artificial color shows the green pepper but not the snow peas and not the carrots.

We present a silicon retina fabricated in standard CMOS 0.6-µm technology. The goal of the sensor is to determine whether or not two images are similar. An image known as the reference image is first used during a programming phase to classify the pixels into two zones that correspond to, respectively, the bright and dark pixels of the reference image. Next, an image is analyzed and the values of the pixels of each zone are summed to produce two signals denoted by Sn and Sb at the outputs of the circuit. If the image under analysis is different from the reference image, then the values of these two signals will also be different from those obtained with the reference image. Our circuit thus implements a pattern matching operation that allows us to determine the similarity between two images using the Sb/Sn plane. The architecture of the sensor is described, and both the simulation and the experimental results are given. Moreover, our pattern-matching operator is compared to the normalized correlation operator commonly used in pattern matching, and its performance is discussed. Finally, we present an example of the application of the sensor.

Some systems are developed to help psychiatrists perform diagnoses. Among the approaches to characterize disease by systems tried, the study of face movements seems very informative concerning the state of the patient. We present a set of methods that enable us to characterize 3-D head and eyes movements in the following conditions: mobility of the subject in a sitting position and background lighting of the room. Our approach yields in a simple way to localize the iris and the characterization of its movement in 3-D space. The 3-D data are derived from stereoscopic system. The absolute 3-D movement of eyeballs and their relative movement with regard to the head are obtained, even when moving. A multiresolution method to deal with movements of great differences in amplitude, such as eye closure, is also presented. No specific lighting is required.

The basic blur distribution approach considered in the restoration of degraded images uses a Gaussian structure to cover a large class of possible distributions. In contrast to this general hypothesis, a common problem encountered in restoration is that the restoration algorithm, especially working in real time, might have large errors with unexpected results if the distribution of image blur based on different media such as atmosphere, x rays, etc. is not known. From this point of view, information about the distribution of blur operating on image is an essential step toward a better restoration to obtain a high-quality image from the degraded one. We introduce a novel approach for restoration of images by using edge and distribution information of the degraded image. For that purpose, first, the distribution model of the degradation process is estimated by a correlation between actual image and reference distribution test models. Then, the proposed algorithm estimates restoration filter model coefficients according to the estimated distribution model and edge information of the degraded image. Simulation results of image restoration illustrate the performance of the proposed method.

We propose a technique for the estimation of the regularization parameter for image resolution enhancement (superresolution) based on the assumptions that it should be a function of the regularized noise power of the data and that its choice should yield a convex functional whose minimization would give the desired high-resolution image. The regularization parameter acts adaptively to determine the trade-off between fidelity to the received data and prior information about the image. Experimental results are presented and conclusions are drawn.

Over the past century a number of factors have been observed to improve the quality of sampled images. Three of these phenomena are reviewed. The reasons for improvement based on perception studies are given, the improvements are quantified, and how they can be used in a practical television system is shown. Interlaced scanning reduces vertical resolution for several reasons involving spatial and temporal masking effects in visual perception. A progressive scan avoids these limitations. Because of the oblique effect in vision and the statistical orientation of lines in scenes, diagonal sampling reduces the required number of pixels in an image. Quantitatively, our measurements show that the number of pixels is reduced by a factor of 1.4 for the same perceived sharpness. In addition, by taking advantage of the octave-wide tuning bands in visual perception, our measurements show that the perceived resolution in the vertical direction for a progressive scan can be double that of an interlaced scan. Using diagonal sampling, a 1920×1080 image with progressive scan at 60 frames/s requires the same transmission bit rate as a 1920×1080 Cartesian sampled image scanned interlaced at 30 frames/s. This results in an image that appears to be much sharper than the 1080 line interlaced format without the interlace artifacts.

We present a robust region segmentation method based on a pseudo-distance map (PDM) that uses a watershed algorithm as a segmentation tool. The PDM is a regularized version of a Euclidean distance map (EDM) directly computed from the edge-strength function (ESF) of an input image without edge detection, which involves a thresholding operation. This unavoidably causes useful region boundary information loss from the original image. We show that applying the watershed algorithm to the PDM significantly reduces oversegmentation, and the final segmentation results obtained by a simple region-merging process are more accurate and meaningful and less sensitive to noise than those of the gradient-based or EDM-based methods. We also propose a simple and efficient region-merging criterion that considers both boundary strengths and inner intensities of regions to be merged. We tested and verified the robustness of our method with a variety of synthetic and real images.

We present a computer vision system for measuring the weight of gobs during a glass-forming process, and a control strategy to automatically correct any weight deviation from a given set point. The system is based on a reliable gob area estimation using image-processing algorithms. A monochrome CCD high-resolution camera and a photodetector for synchronizing acquisition are used for registering gob images. Assuming that the gob has symmetry of revolution about the vertical axis, the proposed system estimates the weight of gobs with accuracy better than ±0.75%. A learning weight control strategy is proposed based on a proportional-integral (PI)-repetitive control scheme. The weight deviation from a set point is used as a control signal to adjust the glass flow into the feeder. This regulation scheme enables effective weight control, canceling mid- and long-term effects. The tracking error of ±1.5% means a reduction of 40% when compared with a traditional PI controller.

A standard approach to extending gray-scale filters to color is the use of gray-scale filters on each of the red, green, and blue bands separately. This RGB component method, when applied to connected filters, causes the introduction of new colors into the image, which could lead to incorrect classification or analysis. We develop a method of applying connected filters to color images using rotation in the hue band of the hue-saturation-value (HSV) color space. The rotation shifts the discontinuity in hue such that the distance to the nearest k-means cluster mean is maximized and the histogram value is simultaneously minimized. When applied to the sample image set, the error from the RGB component approach is of the order of 10 times greater than that of our hue rotation approach. Additionally, our approach yields qualitatively improved images, with fewer new colors introduced than the RGB component method.

A 3-D image-to-device gamut mapping algorithm (I-D GMA) is an ideal way to map display image to the inside of printer gamut. A quick decision whether or not each pixel is located inside of the device gamut is necessary to execute a 3-D I-D GMA. We propose a new gamut boundary descriptor (GBD) for comparing the gamut between image and device by the discrete polar angle (Δθ, Δφ). The gamut shell is described as the simple radial distances, called an r-image. We discuss the location of optimum focal points as mapping centers in an r-image for 3-D I-D GMA. The performance of a 3-D I-D GMA using an r-image is evaluated in the two cases of single and multiple focal points. To maintain the continuous gradation of an image after mapping, the device gamut shell should have a smooth and a precise surface. The printer gamut shell is simply shaped like a polygon made of a triangular mesh pattern, but shaped to have a smoother surface by introducing Overhauser spline functions. Psychophysical experiments were carried out to assess the color appearance matching between the original CRT image and the printed hard copies after mapping. The performance of the 3-D I-D GMA is compared with that of a conventional device-to-device gamut mapping algorithm (D-D GMA). We claim that the 3-D I-D GMA using multiple focal points and a spline GBD achieves the best mapping results for the wide gamut images.

Multiscale error diffusion is superior to conventional error diffusion methods in digital halftoning as it can eliminate directional hysteresis completely. However, there is a bias to favor a particular type of dots in the course of the halftoning process. A new multiscale error diffusion method is proposed to improve the diffusion performance by reducing the aforementioned bias. The proposed method can eliminate the pattern noise in flat regions and the boundary effect found in some other conventional multiscale error diffusion methods. At the same time, it can preserve the local features of the input image in the output. This is critical to quality, especially when the resolution of the output is limited by the physical constraints of the display unit.

A scheme with three key components including wavelet tree coding, error resilient entropy coding (EREC), and error concealment is proposed for robust image coding and transmission over noisy channels. First, we individually encode the spatial-orientation trees in the wavelet domain using the algorithm of set partitioning in hierarchical trees (SPIHT). Error propagation is thus limited because multiple independent bit streams are generated. Meanwhile, a high source coding efficiency is also preserved because the self-similarity property in each wavelet tree remains intact. Next, we use EREC to reorganize these variable-length bit streams into fixed-length data slots before multiplexing and transmission. Therefore, the synchronization of the start of each bit stream can be automatically obtained at the receiver. Finally, to alleviate the possible catastrophic image degradation that may result from errors in the beginning of the bit streams, we propose an error concealment technique to constrain the EREC decoding as well as to postprocess the decoded wavelet coefficients. As a result of the error concealment, the EREC decoding complexity is reduced and the reconstructed image quality is significantly improved. Experimental results demonstrate an excellent error resilient performance of the proposed scheme.

Extended visual cryptography [Ateniese et al., Theor. Comput. Sci.250, 143–161 (2001)] is a method that encodes a number of images so that when the images are superimposed, a hidden image appears while the original images disappear. The decryption is done directly by human eyes without cryptographic calculations. Our proposed system takes three natural images as input and generates two images that are modifications of two of the input pictures. The third picture is viewed by superimposing the two output images. A trade-off exists between the number of gray levels and the difficulty in stacking the two sheets. Our new approach enhances the registration tolerance to obtain the third image and reduces the difficulty of superimposing the image while allowing a variety of gray levels. It is done by extending dot-clustered subpixel arrangements and enabling continuous gray-scale subpixel values. The system has considerably enhanced tolerance to the registration error. We show this by superimposing the output by computer simulation and calculating the peak SNRs with the original images.

This PDF file contains the editorial “Colour Engineering: Achieving Device Independent Colour” for JEI Vol. 13 Issue 03