Comparative performance analysis of spectral estimation algorithms and computational optimization of a multispectral imaging system for print inspection

We have analyzed the performance of simulated multispectral systems for the spectral recovery of reflectance of printer inks from camera responses, including noise. To estimate reflectance we compared the performance of four algorithms which were not comparatively tested using the same data sets before. The criteria for selection of the algorithms were robustness against noise, amount of data needed as inputs (training set, spectral responsivities) and lacking of use of dimensionality reduction techniques. Three different sensor sets and training sets were used. We analyzed the differences in the spanning of the subspaces found for the three training sets, concluding that the ink reflectances have characteristic features. The best performance was obtained using the kernel and the radial basis function neural‐net‐based algorithms for the training set composed of printer inks reflectances, whereas for the other two training sets (composed of samples from the ColorChecker DC and Vhrel's reflectances' set) the quality of the recovered samples was more uniform among the algorithms. We also have performed an optimization to choose the best sensor set for the multispectral system with a reduced number of sensors. © 2012 Wiley Periodicals, Inc. Col Res Appl, 39, 16–27, 2014     

A multispectral imaging approach to colour measurement and colour matching of single yarns without winding

This paper investigates a multispectral imaging approach to colour measurement and colour matching of single yarns. The small size of a single yarn makes it impossible for spectrophotometers directly to acquire its spectral reflectance. Multispectral imaging systems, on the other hand, have the potential to measure the reflectance of single yarns as they can record both the spectral and the spatial information of a sample. A multispectral imaging system, namely imaging colour measurement, has been developed to conduct colour measurement of single yarns. A single yarn is first detected from backgrounds by a modified K‐means clustering method. The reflectance of the single yarn is then specified by an averaging method. Comparative experiments based on 100 pairs of single yarns and corresponding yarn windings show that the reflectance magnitude of a single yarn acquired by imaging colour measurement is smaller than that of corresponding yarn winding measured by a Datacolor 650 spectrophotometer. Experiments on 16 single yarns show that the repeatability and spatial reproducibility of the imaging colour measurement system in measuring a single yarn colour are 0.1185 and 0.2827 CMC(2:1) units. A colour matching comparison experiment (pass or fail), using 24 pairs of single yarns and corresponding pairs of solid‐colour yarn dyed fabrics, shows that single yarns measured by imaging colour measurement can achieve similar colour matching results to solid‐colour yarn dyed fabrics measured by the Datacolor 650 spectrophotometer, with degrees of similarity of 87.5 and 83.3% when the CMC(2:1) and CIE2000(2:1:1) colour difference formulas are employed.     

Comparison of real colour gamuts using a new reflectance database

A large set of data, comprising the spectral reflectances of real surface colours, has been accumulated. The data comprise 16 groups with different materials and include 85,879 measured spectra. From these data, CIELAB colorimetric coordinates were calculated under CIE illuminant D50 and the CIE 1931 standard colorimetric (2°) observer. Several published colour gamuts including those developed by Pointer and ISO reference colour gamut [ISO Graphic Technology Standard 12640‐3:2007] were compared using the present data set. It was found that the Pointer gamut is smaller than the new real data in most of the colour regions. The results also showed that the ISO reference colour gamut is larger than the new real accumulated data in most regions. The present finding indicates that there is a need to derive a new colour gamut based on the newly accumulated data for common applications. © 2013 Wiley Periodicals, Inc. Col Res Appl, 39, 442–451, 2014     

Artwork imaging from 370 to 1630 nm using a novel multispectral system based on light‐emitting diodes

The recent use of multispectral systems as a noncontact method for analysis of artworks has already shown promising results. This study explains the application of a novel portable multispectral system based on light‐emitting diodes (LEDs) for artwork imaging. This method provides spectral information in a spectral range from 370 to 1630 nm with a 25 cm × 25 cm field of view by using two different image sensors in synchrony with 23 bands of irradiation. The spectral information for each point is estimated and validated using the pseudo‐inverse and spline interpolation methods for spectral estimation and three different evaluation metrics. The results of the metrics obtained with both estimation methods show a general good performance of the system over the whole spectral range. The experiments also showed that the selection of the training set for the pseudo‐inverse estimation has a great influence in its performance, and thus, it defines whether or not the pseudo‐inverse outperforms the spline interpolation method. The system is applied in situ to the study of Catalan art masterpieces, and the results demonstrate the potential of a cost‐effective and versatile system using various off‐the‐shelf elements to reconstruct color information and to reveal features not previously identified. © 2014 Wiley Periodicals, Inc. Col Res Appl, 40, 398–407, 2015     

Color measurement of single yarn based on hyperspectral imaging system

This article proposed a novel approach to color measurement of a single yarn using hyperspectral imaging system (HIS). Due to the size of a single yarn, it is impossible for spectrophotometers to measure its color directly. The HIS can acquire the spectral reflectance of continuous bands within a region of interest on a yarn sample, which can achieve color measurement of a single yarn compared with traditional spectrophotometers. A single yarn is segmented from the background by a spectral matching method through adaptively setting threshold of Fréchet distance values. The spectral reflectance of single yarn is specified by a method that lightness of pixels used as weight. The experiment based on Pantone Cotton Chip Set shows that the interinstrument agreement between the HIS and a standard spectrophotometer Datacolor SF650 has a significant improvement after using the R-Model, and the average percentage improvement of the color difference is up to 54.99%. The yarn segmentation comparative experimental results show that the proposed method to segment single yarn from background is better in retaining the edge information of the yarn than the modified K-means clustering method, and the color of the yarn segmented by the proposed method is more similar to the actual color of single yarn.     

Spectral reflectance reconstruction from RGB images based on weighting smaller color difference group

A method to reconstruct spectral reflectance from RGB images is presented without priori knowledge of camera's spectral responsivity. To obtain the spectral reflectance of a pixel or region in images, this method assumes that reflectance is a weighted average of reflectances of samples in a selected training group, in which all samples have smaller color difference with that pixel or region. Four proposed weighting modes with different selected numbers of training samples were investigated. Among them, the inverse square weighting mode obtains the best performance, and it is not very sensitive to the selected training samples number. Experimental results show that all weighting modes outperform the traditional method in terms of root mean squared error and Goodness‐of‐Fit Coefficient between the actual and the reconstructed reflectances as well as color differences under the other light condition. © 2016 Wiley Periodicals, Inc. Col Res Appl, 42, 327–332, 2017     

Recovery and representation of spectral bidirectional reflectance distribution function from an image‐based measurement system

A spectral‐based method can acquire and represent the surface appearance of a given material physically correctly. But, it has drawbacks due to its high measurement cost and a long computation time in measuring, modeling, and rendering. In this article, we present spectral recovery and representation of spectral bidirectional reflectance distribution function (BRDF) from multispectral reflectance measurements in which we can render real appearance materials over a 3D model with accuracy and efficiency. First of all, an accurate spectral BRDF recovery algorithm, which transforms multispectral high dynamic range images into highly dense BRDFs in both a spectral and an angular domain, is proposed. Second, an efficient representation method is developed representing spectral BRDFs compactly using a factorization method and an adaptive spectral sampling method that uses a given error bound. The results show that the proposed method can compress the spectral BRDF data down by several hundred times while maintaining the given accuracy in colorimetric and spectral domains under a specific illuminant. © 2015 Wiley Periodicals, Inc. Col Res Appl, 41, 358–371, 2016     

Colour and spectral reflectance of stools from normal neonatal babies

Data is presented on the colour and spectral reflectances of stool samples from healthy young infants (mean age 6 days). The range of spectral reflectances is well described by a pigment mixture model with three components—a pale component corresponding to partially digested milk fats, a yellow component corresponding to bilirubin breakdown products, and a dark component corresponding to meconium and bilirubin. Most stools were tan, beige or olive but white, grey and light grey were also observed. © 2014 Wiley Periodicals, Inc. Col Res Appl, 40, 585–591, 2015     

Spectral dependence of colorimetric characterisation of scanners

The spectral dependence of the colorimetric characterisation of a typical scanner was investigated. Different colour sets, including ColorChecker SG, Kodak Q‐60 colour input target, a set of plain woven coloured fabrics with a large colour gamut, and randomly selected samples of the Munsell Book of Color were used as training and testing sets in the colorimetric characterisation of a scanner by employing a non‐linear regression method. The coefficient matrices were optimised for each particular media in the training stage and used to predict the device‐independent colorimetric data, i.e. CIELab values of other media from their corresponding RGB values measured by the scanner. In order to extract the differences between the applied sets and determine the actual dimensions of their reflectance spectra, the principal component analysis technique was employed. As expected, it was observed that the different sets benefit from diverse dimensional properties and, in some cases, the spectral behaviours of the first few eigenvectors were apparently different. It was demonstrated that scanner colorimetric characterisation depended on the spectral properties of the applied colour set in the training stage and, consequently, the testing errors increased with increasing the spectral dissimilarity between the sets that were used in training and testing sequences. It was concluded that, to achieve better colour reproduction results, the scanner should be characterised for each media with specific spectral properties.     

An efficient method for solid‐colour and multicolour region segmentation in real yarn‐dyed fabric images

This paper presents an efficient approach to solid‐colour and multicolour region segmentation in real yarn‐dyed fabric images. The approach is based on a novel model describing the spectral response of a multispectral imaging system to yarn‐dyed fabrics. The model indicates that solid‐colour regions cannot be distinguished from multicolour regions in terms of reflectance, tristimulus, or CIELAB values owing to a geometric term representing the influence of fabric surface condition on measured colours. The geometric term makes it difficult to determine the segmentation thresholds of CIEXYZ and CIELAB colour histograms. However, solid‐colour and multicolour regions can be detected in CIExyY space because chromaticity coordinates are impervious to the geometric term. The CIExyY histograms of a solid‐colour region accord with one Gaussian distribution, but those of a multicolour region accord with a combination of two Gaussian distributions. The CIEXYZ, CIELAB, and CIExyY colour distributions of both solid‐colour and multicolour yarn‐dyed fabrics were analysed in detail in simulation and real experiments. Experimental results show that solid‐colour yarn‐dyed regions can be distinguished from multicolour yarn‐dyed fabric regions by the shapes of CIExyY histograms, but cannot be distinguished by the shapes of CIEXYZ or CIELAB histograms.     

Improving unsupervised saliency detection by migrating from RGB to multispectral images

Saliency detection has been an important topic during the last decade. The main goal of saliency detection models is to detect the most relevant objects in a given scene. Most of these models use RGB (Red, Green, Blue) images as an input because they mainly focus on applications where features (eg, faces, textures, colors, or human silhouettes) are extracted from color images, and there are many labeled databases available for RGB‐based saliency data. Nevertheless, the use of RGB inputs clearly limits the amount of information from where to extract the salient regions as spectral information is lost during the color image recording. On the contrary, multispectral systems are able to capture more than three bands in a single capture and can retrieve information from the full spectrum at a pixel. The main aim of this study is to investigate the advantages of using multispectral images instead of RGB images for saliency detection within the framework of unsupervised models. We compare the performance of several unsupervised saliency models with both RGB and multispectral images using a specific dataset of multispectral images with ground‐truth data extracted from observers' fixation patterns. Our results show a general improvement when multispectral information is taken into account. The saliency maps estimated by using the multispectral features are closer to the ground‐truth data, with the simplest Graph‐based visual saliency and Boolean Map‐based models showing good relative gain compared with other approaches.     

Colorimetric calibration of a video digitizing system: Algorithm and applications

In principle, digitized color video images should be rich and convenient sources of colorimetric information. In practice, these advantages are offset by the difficulty of reliably translating the video camera's output into colorimetric variables. A solution to this problem is outlined here, one which exploits the fact that spectral reflectances of many natural materials vary slowly in the visible. A characteristic vector analysis of reflectances for a set of such materials leads to an algorithm that gives colorimetrically accurate spectral reflectances from the red-green-blue output of a video digitizing system. Prior knowledge about the illumination leads to chromaticity and luminance information, which can be comparable in quality to that obtained from a spectrora-diometer. Some sample retrievals are shown for the algorithm. Since it is designed to correct color biases that are unknown initially, the algorithm has the advantage that images from many sources can be analyzed.     

Recovering spectral data from natural scenes with an RGB digital camera and colored filters

Many spectral‐recovery methods using RGB digital cameras assume the underlying smoothness of illuminant and reflectance spectra, and apply low‐dimensional linear models. The aim of the present work was to test whether a direct‐mapping method could be used instead of a linear‐models approach to recover spectral radiances and reflectances from natural scenes with an RGB digital camera and colored filters. In computer simulations, a conventional RGB digital camera with up to three colored filters was used to image scenes drawn from a hyperspectral image database. Three measures were used to evaluate recovery with the direct‐mapping method: goodness‐of‐fit, root‐mean‐square error, and a color‐difference metric. It was found that with two and three filters both spectral radiances and reflectances could be recovered sufficiently accurately for many practical applications. With little increase in computational complexity, an RGB camera and a few colored filters can provide significantly better recovery of natural scenes than an RGB camera alone. © 2007 Wiley Periodicals, Inc. Col Res Appl, 32, 352–360, 2007     

A novel method for weft and warp yarn segmentation in multicolour yarn‐dyed fabric images

This paper proposes a novel method for segmentation of weft and warp yarns in multicolour yarn‐dyed fabric images. A multicolour yarn‐dyed fabric is cross‐woven by weft and warp yarns with different colours. When a multispectral imaging system is used to measure the colour of a multicolour yarn‐dyed fabric image, its weft and warp yarns need to be detected before analysing their colours. Detection of interstices between weft and warp yarns is firstly conducted. A modified K‐means clustering approach is then utilised to separate weft and warp yarns. The number of clusters is fixed to 2. The metric to measure the distance between a pixel and the mean of a cluster is the CIELAB colour difference. The initial means are determined by the expected values of fitted Gaussian distributions to CIExyY colour histograms. Experimental results show that the proposed method is promising for the segmentation of weft and warp yarns in multicolour yarn‐dyed fabrics, with an improved segmentation accuracy and much faster processing speed than K‐means clustering in CIEXYZ and CIELAB spaces.     

Effective models for correlating spectral reflectance between spectrophotometers

This paper describes different models to achieve satisfactory interinstrument agreement between spectrophometers. The models investigated include the Berns and Petersen method, the modified Berns and Petersen method, the Reniff method, the Rich and Martin method and the R‐model developed by Chung et al. In the paper, the Reniff method and the R‐model were extended, and all the models were then evaluated using sample sets with different substrates. It was found that the extended Reniff method gave the best training performance and the worst testing performance. The Rich and Martin model, the three‐term Chung model and the Berns and Petersen models, all with wavelength‐dependent coefficients, gave better testing performance.     

Recovering spectral information using digital camera systems

There is an increasing need to be able to measure colour properties of complex surfaces or images for which traditional spectrophotometers are not suitable. New multispectral imaging systems are being developed but it is not clearly understood how the parameters (such as the number of colour channels, the spectral properties of the channels, and the choice of illuminant) of such systems affect the performance. Furthermore, the effect of sensor and quantisation noise on the overall performance of the system also needs to be considered. This paper describes the development of a mathematical model of a multispectral imaging system that takes into account imaging parameters and noise. The results from the computational model show that increasing the number of colour channels alone in the imaging system does not necessarily allow better estimates of spectral reflectance. The choice of illumination can also, in the presence of noise, greatly affect performance.     

The minimum number of measurements for colour,sparkle, and graininess characterisation in gonio‐apparent panels

Materials with new visual appearances have emerged over the last few years. In the automotive industry in particular there is a growing interest in materials with new effect finishes, such as metallic, pearlescent, sparkle, and graininess effects. Typically, for solid colours the mean of three measurements with repetitions is sufficient to obtain a representative measurement for colour characterisation. However, gonio‐apparent panels have non‐homogeneous colours, and there are no studies that recommend the minimum number of repetitions for colour, sparkle, and graininess characterisation of this type of panel. We assume that colour panels incorporating special‐effect pigments in their colour recipes will require a higher minimum number of measurements than solid colour panels. Therefore, the purpose of this study is to verify this assumption by using a multiangle BYK‐mac spectrophotometer, given that it is currently the only commercial device that can measure colour, sparkle, and graininess values simultaneously. In addition, a possible methodology is given for establishing the minimum number of measurements when characterising gonio‐apparent materials using a specific instrument, able to be implemented in future instruments when determining multiple appearance attributes (colour, gloss, sparkle, etc.) for many coloration technologies. Thus, we studied the minimum number of measurements needed to characterise the colour, sparkle, and graininess of three types of sample with solid, metallic, and pearlescent coatings respectively. Twenty measurements were made at twenty random positions (different target areas) of 90 samples. The minimum number of measurements for all these variables was determined on the basis of the point at which the cumulative mean value became constant. Thus, applying new statistical tools, it is clearly shown that metallic and pearlescent panels require more colour measurements than solid panels, in particular when geometries are being measured in a specular direction. As regards texture (sparkle and graininess), more measurements are needed for graininess than for sparkle, and more for metallic panels than for pearlescent panels.     

A custom ink‐jet printing system using a novel pretreatment method

A low‐cost and environmental‐friendly direct dye‐based ink‐jet printing system was developed. A novel ink‐jet pretreatment method was employed, in which the cationic fixing agent, Matexil FC‐ER, was applied as the colourless fixing ink and applied only on image areas of the fabric by ink‐jet printer. It was found that this new pretreatment method could more effectively enhance the colour strength and improve the wash fastness (greyscale ≥ 3) when compared with traditional exhaust application. The cross‐staining of non‐image areas of fabric was also apparently decreased using this new method. The light fastness of ink‐jet pretreated samples was slightly reduced as the presence of Matexil FC‐ER made dyes more sensitive to light.     

Surface reflectance estimation by the dichromatic model

A method is described for estimating the surface-spectral reflectances of glossy objects when the color signal is a mixture of diffuse reflections, specular reflections, and interreflections. The objects are inhomogeneous dielectric materials, and the reflected light is measured using a spectroradiometer. We first describe the main idea; the color signals, reflected from two closely apposed surfaces with a single interreflection between them, can be expressed by a linear combination of the illuminant spectrum and two diffuse spectral reflection functions. We introduce a representation in which each of these three terms is projected onto a point on a unit sphere. Estimation of the diffuse reflection functions is then reduced to finding the vertices of the spherical triangle. Next, an algorithm is described to estimate the locations of the vertices of diffuse reflectance functions from the measured samples. The reliability of the algorithm is demonstrated in an experiment using two plastic objects with glossy surfaces. © 1996 John Wiley & Sons, Inc.     

Mathematical approach for predicting non‐negative tristimulus values using the CAT02 chromatic adaptation transform

It has been reported that for certain colour samples, the chromatic adaptation transform CAT02 imbedded in the CIECAM02 colour appearance model predicts corresponding colours with negative tristimulus values (TSVs), which can cause problems in certain applications. To overcome this problem, a mathematical approach is proposed for modifying CAT02. This approach combines a non‐negativity constraint for the TSVs of corresponding colours with the minimization of the colour differences between those values for the corresponding colours obtained by visual observations and the TSVs of the corresponding colours predicted by the model, which is a constrained non‐linear optimization problem. By solving the non‐linear optimization problem, a new matrix is found. The performance of the CAT02 transform with various matrices including the original CAT02 matrix, and the new matrix are tested using visual datasets and the optimum colours. Test results show that the CAT02 with the new matrix predicted corresponding colours without negative TSVs for all optimum colours and the colour matching functions of the two CIE standard observers under the test illuminants considered. However, the accuracy with the new matrix for predicting the visual data is approximately 1 CIELAB colour difference unit worse compared with the original CAT02. This indicates that accuracy has to be sacrificed to achieve the non‐negativity constraint for the TSVs of the corresponding colours. © 2011 Wiley Periodicals, Inc. Col Res Appl, 2011