Image and Signal Processing

In this study we introduce three daylight colored spectra, i.e. spectra with correlated color temperatures near 6500K, for improved color discrimination. This property has been estimated by the volume of the object color solid in a nearly uniform color space based on the DIN99d color difference formula. Three optimized spectra produce about 11% - 13% larger volume than the standard D65 illuminant which simulates natural daylight and improve especially the red-green color discrimination. The optimal spectra are the result of similar optimization processes, but differ in shapes, except the common gap in light power in the region 570 nm - 610 nm.

In this research, we evaluated the visibility of age spot and freckle by changing volume of pigmentations and the spatial distribution of age spot and freckle based on the spectral reflectance of skin. The spectral reflectance is simulated by using Monte Carlo simulation of light transport in multi-layered tissue. Three type of spatial distributions of age spot and freckle are generated based on the simulated spectral reflectance. We performed subjective evaluation for the visibility of the simulated age spot and freckle patterns, and found age spot and freckle become less noticeable as the increase of blood volume.

A color matching experiment was designed and carried out to estimate human observers’ color matching functions (CMFs). 61 color-normal observers participated in the experiment. Their results were traced back to physiological factors using a mathematical vision model. The experiment time was 15 minutes, which is much faster than previous research aimed at determining color matching functions.

Liquid crystal tunable filters can provide rapid and vibrationless section of any wavelength in transmitting spectrum so that they have been broadly used in building multispectral or hyperspectral imaging systems. However, the spectral range of the filters is limited to a certain range, such as visible or near-infrared spectrum. In general hyperspectral imaging applications, we are therefore forced to choose a certain range of target spectrum, either visible or near-infrared for instance. Owing to the nature of polarizing optical elements, imaging systems combined with multiple tunable filters have been rarely practiced. In this paper, we therefore present our experience of building a two-way hyperspectral imaging system with liquid crystal tunable filters. The system allows us to capture hyperspectral radiance continuously from visible to near-infrared spectrum (400—1100 nm at 7 nm intervals), which is 2.3 times wider and 34 times more channels compared to a common RGB camera. We report how we handle the multiple polarizing elements to extend the spectral range of the imager with the multiple tunable filters and propose an affine-based method to register the hyperspectral image channels of each wavelength.

A method is proposed for measurement and modeling of bidirectional characteristics of fluorescent objects. First, a gonio-spectro measurement system is constructed for measuring the spectral luminescent radiance factor of a variety of fluorescent object surfaces. Second, the angular dependency of the luminescence radiance factor is analyzed in different light incidence and viewing directions. The observed radiance factors can then be described by the Lambertian model with good accuracy. We also analyze the bidirectional reflection radiance factor of a white mat surface. The whole characteristics of bispectral bidirectional radiance factors of a fluorescent object can be summarized as a compact mathematical model. Finally, image rendering of a fluorescent object is performed using the Donaldson matrix estimated in a separate measurement system. The feasibility of the proposed method was examined experimentally.

Precancerous lesions are in many situations not visible during white light gastroendoscopy. Different approaches have been proposed based on light tissue interaction in order to improve the visualization by creating false color images. However, these systems are limited to few wavelengths. In this paper, we propose a multispectral gastroendoscopic system and a methodology to identify precancerous lesions. The multispectral images collected during gastroendoscopy are used to compute statistical features from their spectrum. Pooled variance t-test is used to rank the features in order to train 3 classifiers with different number of features. The 3 classifiers are Neural Networks using Generalized Relevance Learning Vector Quantization (GRLVQ), SVM with a Gaussian kernel and K-nn. The performance is compared based on their ability to identify precancerous lesions, using as quantitative index the accuracy, specificity and sensitivity. SVM presents the best performance, showing the effectiveness of the method.

Spectral imaging is a useful tool in many fields of scientific research and industry. Spectral images contain both spatial and spectral information of the scene. Spectral information can be used for effective visualization of the features-of-interest. One approach is to use spectral image enhancement techniques to improve the diagnostic accuracy of medical image technologies like retinal imaging. In this paper, two multichannel spectral image enhancement methods and a technique to further improve the visualization are presented. The methods are tested on four multispectral retinal images which contain diabetic retinopathy lesions. Both of the methods improved the detectability and quantitative contrast of the diabetic lesions when compared to standard color images and are potentially valuable for clinicians and automated image analyses.

LED illumination based multispectral imaging (LEDMSI) is one of the promising techniques of fast and effective spectral image acquisition. Several LEDMSI systems and methodologies have been proposed in the literature. A typical LEDMSI system uses a monochrome camera, which captures images of a scene under

n

different color LED lights, producing an

n

-band spectral image of the scene. RGB camera based LEDMSI systems have been proposed to speed up the acquisition process. However, demosaicing process in these systems affects the spatial accuracy, and in turn influences the quality of resulting spectral images. In this paper, we study how the performance and quality of LEDMSI systems are influenced by different factors. Four major factors: camera type, demosaicing, number of color LEDs and, noise are considered in the study. We carry out simulation experiments using monochrome and RGB camera based LEDMSI systems, under the influence of different amounts of noise and practical constraints on the number of different color LEDs. The experiments confirm the influence of these factors on the performance of a LEDMSI system. We believe that this work would be useful not only in designing LEDMSI systems, but also in developing quality framework(s) for the evaluation of spectral images and spectral imaging systems.

In gastrointestinal surgery, surgeon subjectively judges if the organ is healthy from the color. However it is difficult to discriminate a small difference of organ’s color by visual inspection. In this paper, we focus on the tissue oxygen saturation (StO

2

) that represents balance of oxygen demand and supply in tissue and try to estimate its value by transmitted light intensity analysis. We developed a system for measurement of transmitted light intensity using a compact spectrometer and a halogen light source and collected transmitted light intensity data from pig’s small intestines. Absorbance of the tissue was then calculated from those data. On the basis of Beer-Lambert law, we estimated StO

2

from the calculated absorbance. Results of evaluation experiment to pig’s small intestines suggested the possibility of quantitative evaluation of tissue viability by the proposed method.

Density measurement of printed color samples takes an important role in print quality inspection and process control. When multi-spectral imaging systems are considered for surface reflectance measurement, the possibility of calculating spectral print density over the spatial image domain arises. A drawback in using multi-spectral imaging systems is that some spectral reconstruction algorithms can produce estimated reflectances which contain negative values that are physically not meaningful. When spectral density calculations are considered, the results are erroneous and calculations might even fail in the worst case. We demonstrate how this problem can be avoided by using kernel ridge regression with additional link functions to constrain the estimates to positive values.

The existing spectral colour difference metrics are not similar to CIEDE2000. The goal in this study was to implement a system to calculate the difference of spectral colours so that the calculated differences are similar to CIEDE2000 colour differences. The developed system is based on a priori calculated differences between known spectra and the calculus parameters derived from them. With the current system one can calculate spectral differences between a limited set of spectra which are derived by mixing the known spectra. The computation of calculus parameters for the system is a demanding process, and therefore, the calculations were distributed to a cluster of computers. The proposed spectral difference metric is very similar to CIEDE2000 for most of the test spectra. In addition, the metric shows non-zero differences for metameric spectra although CIEDE2000 colour difference metric results in zero differences. This indicates more correct operation of the spectral difference than the operation of CIEDE2000 colour difference.

In this study spectral and spatial characterization of a seven channel FluxData 1665 MS7 three-CCD spectral video camera was performed in terms of the sensor spectral sensitivity, linearity, spatial uniformity, noise and spatial alignment. The results indicate small deviation from ideal linear sensor response. Also, a small spatial non-uniformity and systematic shift exists between the channel images. However, images were observed to have high quality in term of noise. Spectral characterization showed that the sensor has good response in the 380-910 nm region with only some sensitivity limitations in the 715-740 nm range. We also evaluated the spectral reflectance estimation in 400-700 nm range using empirical regression methods and the Digital ColorChecker SG and ColorChecker charts. These experiments resulted in average Δ

E

00 color accuracy of 1.6 – 2.4 units, depending on the illuminant and estimation method.

Poisson distributed noise, such as photon noise is an important noise source in multi- and hyperspectral images. We propose a variational based denoising approach, that accounts the vectorial structure of a spectral image cube, as well as the poisson distributed noise. For this aim, we extend an approach for monochromatic images, by a regularisation term, that is spectrally and spatially adaptive and preserves edges. In order to take the high computational complexity into account, we derive a Split Bregman optimisation for the proposed model. The results show the advantages of the proposed approach compared to a marginal approach on synthetic and real data.

The technological fields where solid-state lighting can be applied are constantly growing. In relation to this topic, we built a spectral LED-based tuneable light source for the reconstruction of CIE standard illuminants. This light source consists of 31 spectral channels from 400nm to 700nm, an integrating cube and a control board with 16 bit resolution. Moreover, a minimization routine was developed to calculate the weighting values per channel for reproducing standard illuminants. The differences in colorimetric and fitting parameters between standard spectra and the theoretical and experimental ones, showed that the reconstructed spectra were very similar to the standard ones, specially for the D65 and A illuminants. However, there was a certain mismatching from 500nm to 600nm due to the lack of LEDs in this region. In conclusion, the developed LED-based light source and minimization routine are able to reproduce CIE standard illuminants with high accuracy.

Many kinds of spectral image formats are used for various applications, but there is still no existing standard format. Natural Vision data file format is one of the best possible candidates for the standard of spectral image format due to its flexibility to adapt to various kinds of existing image format and capacity to include information needed for each application. In biology, the analysis of huge datasets acquired by various techniques requires the use of specific databases. In order to be able to combine different data, defining a standard spectral image format that includes biological parameters is of prime importance. This paper describes an attempt to use Natural Vision data file format for spectral images related to biology and highlights the merits of Natural Vision data file format as an application oriented spectral image format.

This paper experimentally shows how the center wavelength and spectral power distribution (SPD) of displayed color is related to chromostereopsis. Chromostereopsis - a visual illusion whereby the impression of depth is conveyed in two-dimensional color images - can be applied to glassless binocular stereopsis by controlling color saturation even when a commercial liquid crystal display (LCD) is used to display a two-dimensional image. We conducted evaluations of stereoscopic visual effects among monochrome images using an LCD panel and three monochrome backlights whose SPD had a single peak. The center wavelength and full width at half maximum (FWHM) of the SPD for the backlight were varied. The experimental results show that chromostereopsis does not occur strongly when the FWHM of a backlight is larger than 100 nm. We also suggest that the impression of the depth for monochrome images depends on the center wavelength and FWHM of the color, which indicates chromostereopsis can be expressed by the chromatic aberration.

Hybrid-resolution spectral imaging is a technique that efficiently produces high-resolution spectral images by combining low-resolution spectral data with a high-resolution RGB image. In this paper, we introduce a regression- based spectral reconstruction method for this system to enable us doing accurate spectral estimation without a laborious measurement of the spectral sensitivity of the RGB camera. We present two methods for regression-based spectral reconstruction that utilize spatially-registered pair of a low-resolution spectral image and a high-resolution RGB image: whole frame data regression and locally weighted regression. In the experiment, we developed a hybridresolution spectral imaging system, and it was confirmed that the regressionbased methods can estimate spectra in high accuracy.

Linear interpolation methods have the characteristics of low computational complexity which makes them widely developed in CFA (color filter array) demosaicking. However, the trichromatic nature of colour images enables CFA demosaicking algorithms to take advantage of the luminance-chrominance representation to interpolate the colour planes efficiently and effectively. It seems, however, this does not apply to multispectral images in a straightforward manner. In this paper, we first propose a linear interpolation method for SFA (spectral filter array) demosaicking drawing on the mathematical analysis of mosaicking, demosaicking processes and the idea of residual interpolation. We then compare the performance of the proposed method with that of five other techniques by means of the SSIM index. The result shows that our new algorithm has a good performance with less computing time.

In this paper we present an original semi-supervised method for the segmentation of in situ tree color images which combines color quantization, adaptive fragmentation of learning areas defined by the human operator and labeling propagation. A mathematical morphology post-processing is introduced to emphasize the narrow and thin structures which characterize branches. Applied in the L*a*b* color system, this method is well adapted to easily adjust the learning set so that the resultant labeling corresponds to the accuracy achieved by the human operator. The method has been embarked and evaluated on a tablet to help tree professionals in their expertise or diagnosis. The images, acquired and processed with a mobile device, present more or less complex background both in terms of content and lightness, more or less dense foliage and more or less thick branches. Results are good on images with soft lightness without direct sunlight.

This paper proposes a new color representation. This representation belongs to the penta-valued category and it has three chromatic components (

red, blue and green

) and two achromatic components (

black

and

white

). The proposed penta-valued representation is obtained by constructing a fuzzy partition in the RGB color space. In the structure of the penta-valued representation, it is defined the well known negation operator and supplementary, two new unary operators: the dual and the complement. Also, using the Bhattacharyya formula, it is defined a new inter-color similarity. Next, the obtained inter-color similarity is used in the framework of k-means clustering algorithm. On this way, it results a new color image clustering method. Some examples are presented in order to prove the effectiveness of the proposed multi-valued color descriptor.

The luminance of natural scenes frequently presents a high dynamic range and cannot be adequately captured with traditional imaging devices. Additionally, even if the technology to capture the scene is available, the image to be displayed on conventional monitors must be compressed by a tone mapping operator. Exposure fusion is an affordable alternative which blends multiple low dynamic range images, taken by a conventional camera under different exposure levels, generating directly the display image. In this paper, the Retinal-like Sub-sampling Contrast metric has been adapted to work with the original version of the exposure fusion algorithm in the CIELAB color space. In addition, saturation and well-exposedness metrics have been reformulated in this color space, adding a dynamic adjustment mechanism to the latter one which avoids amplification of invisible contrast. Results based on objective evaluation show that the proposed algorithm clearly outperforms the original exposure fusion technique and most of the state-of-the-art tone mapping operators for static images.

A large number of Image Quality (IQ) metrics have been developed over the last decades and the number continues to grow. For development and evaluation of such metrics, IQ databases with reference images, distortions, and perceptual quality data, is very useful. However, existing IQ databases have some drawbacks, making them incapable of evaluating properly all aspects of IQ metrics. The lack of reference image design principles; limited distortion aspects; and uncontrolled viewing conditions. Furthermore, same sets of images are always used for evaluating IQ metrics, so more images are needed. These are some of the reasons why a newly developed IQ database is desired. In this study we propose a new IQ database, Colourlab Image Database: Image Quality (CID:IQ), for which we have proposed methods to design reference images, and different types of distortions have been applied. Another new feature with our database is that we have conducted the perceptual experiments at two viewing distances. The CID:IQ database is available at

http://www.colourlab.no/cid

.

While the RGB color model refers to the biological processing of colors in the human visual system, the HSV color model corresponds to the human perception of color similarity. In this paper we formulate a projection of RGB vectors within the RGB color space, which separates achromatic from chromatic information. The projection is the mathematical equivalent to Hue and Saturation of the HSV color space in the RGB space. It integrates the psycho- visual concept of human differentiation between colors of the HSV space into the physiological-visual based concept of the RGB space. With the projection it is, contrary to the prevailing opinion, possible to differentiate between colors based on human perception in the linear geometry of the RGB color space. This opens new possibilities in many fields of color image processing, especially in the domain of color image segmentation, where color similarity plays a major role.

Much research has been concerned with the contribution of the low level features of a visual scene to the deployment of visual attention. Bottom-up saliency models have been developed to predict the location of gaze according to these features. So far, color besides to brightness, contrast and motion is considered as one of the primary features in computing bottom-up saliency. However, its contribution in guiding eye movements when viewing natural scenes has been debated. We investigated the contribution of color information in a bottom-up visual saliency model. The model efficiency was tested using the experimental data obtained on 45 observers who were eye tracked while freely exploring a large data set of color and grayscale videos. The two datasets of recorded eye positions, for grayscale and color videos, were compared with a luminance-based saliency model [1]. We incorporated chrominance information to the model. Results show that color information improves the performance of the saliency model in predicting eye positions.

In this paper, we explore an original way to compute texture features for color images in a vector process. To do it, we used a dedicated approach for color mathematical morphology using distance function. We show in this paper the scientific construction of morphological spectra and preliminary results using Outex database.

The paper addressed the problem of the script characterization in the old Slavic printed documents. Therefore, an algorithm for the script discrimination was proposed. It was based on the typographical feature classification, which creates ciphers from different scripts of the document. Then, the feature extraction was achieved by statistical analysis. The obtained features were set and stored for further analysis in order to identify the discrimination criteria between different scripts. The proposed method is tested on the example of the Slavic printed documents which contains Glagolitic and Cyrillic script.

3D representations were made of a small Egyptian scarab with a gold band by a number of methods, based on photogrammetry and photometric stereo. They were evaluated for colour fidelity and spatial detail, in the context of a study of toolmarks and manufacturing techniques of jewellery in ancient Egypt. It was found that although a 3D laser scanner gave the best geometric accuracy, the camera-based methods of photogrammetry and photometric stereo gave better representation of fine detail and colour on the object surface.

Hyperspectral imaging is a promising non-invasive method for applications in conservation of painting. With its ability to capture both spatial and spectral information which relates to physical characteristics of materials, the identification of pigments and its spatial distribution across the painting is now possible. In this work, The Scream (1893) by Edvard Munch is acquired using a hyperspectral scanner and the pigment mapping of its constituent pigments are carried out. Two spectral image classification methods, i.e. Spectral Angle Mapper (SAM) and Spectral Correlation Mapper (SCM), and a fully constrained spectral unmixing algorithm combined with linear mixing model are employed for the pigment mapping of the painting.

Our cultural heritage is constituted by irreplaceable artworks that must be known and preserved. Their study and documentation should be in principle carried out using non-invasive approaches. The technological advances in spectroscopic imaging acquisition devices made it possible to apply this methodology to such purpose. In this context, the present paper discusses a particularly challenging task within the conservation field, which is the identification of red lake pigments in artworks, applying Vis-NIR hyper-spectral imaging spectroscopy. The latter was used to characterize and discriminate between historically accurate paint reconstructions of brazilwood (vegetal) and cochineal (animal) lake pigments. The same paints were also analyzed with Fiber Optic Reflectance Spectroscopy to validate the data obtained with the imaging method. The requirements for a successful identification of these pigments are addressed, and future research is suggested in order to increase the usefulness of the technique’s application.

Serbia-Forum is a web application portal designed and implemented by the Mathematical Institute of the Serbian Academy of Sciences and Arts (MISANU) whose goal is to digitally make available, many units of cultural heritage belonging to the national heritage of the republic of Serbia. The Serbia - Forum project as a whole is geared towards the digitization, presentation and organization of digitized Serbian cultural heritage. In the past two years since the onset of the Serbia-Forum project, MISANU in cooperation with many partner institutions both in and out of the Serbian government has enriched its collection of digitized content with 80.000 new units. Serbia-Forum is growing successfully both in impact and in content. Currently, the portal content ranges from postcards, newspapers, photographs, books and other relevant media.

The article reports the research conducted to enhance the readability of Khirbet Qeiyafa ostracon using the spectral image analysis combined with image processing. The spectral imaging carried out in the visible and IR range facilitated the detection, improvement in readability of characters, and better interpretation of the history. Analysis of the spectral data using principal component analysis and independent component analysis showed better ink visibility and gave further support to archaeologists and historians. The proposed techniques resulted in an improvement compared to earlier interpretations.

In the context of an industrial application for securing identity cards, image watermarking is used to verify the integrity and authenticity of images printed on plastic cards support. In this area the watermark must survive image modifications related to print-and-scan process and the degradations submitted by ID cards through its lifetime. In this work various card durability attacks were studied (bending, scratches, color fading) as well as some additional other attacks that reflects other possible attacks in the given industrial context. The ID images were watermarked by a new Fourier watermarking method based on print/scan counterattacks (blurring correction and colors restorations) that allows improvement of the robustness of the watermarking method. Results show a noticeable increase of the overall performances of the Fourier watermarking method against durability attacks that is suitable in the context of the industrial application of interest.

Libraries contain huge amounts of arabic printed historical documents which cannot be available on-line because they do not have a searchable index. The word spotting idea has previously been suggested as a solution to create indexes for such a collecton of documents by matching word images. In this paper we present a word spotting method for arabic printed historical document. We start with word segmentation using run length smoothing algorithm. The description of the features selected to represent the words images is given afterwards. Elastic Dynamic Time Warping is used for matching the features of the two words. This method was tested on the arabic historical printed document database of Moroccan National Library.

Today, handwriting recognition is one of the most challenging tasks and exciting areas of search in computer science. Indeed, despite the growing interest in this field, no satisfactory solution is available. For this reason Multiple Classifier Systems (MCS) based on the combination of outputs of a set of different classifiers have been proposed as a method for the developing of high performance classifier system. In this paper we describe a serial combination scheme of an Arabic Optical Character Recognition System. The classification engine is based on Adaptive Resonance Theory and Radial Basic Function, where an RBF network acting as the first classifier is properly combined with a set of ART1 network (one for each group) trained to classify the word image. The experiments applied on the IFN/ENIT database show that the proposed architecture exhibits best performance.

In this work, we aimed to further improve the commonly used DCTDWT based watermarking method by combining the DCT-DWT with the FABEMD method. Rather than embedding the watermark in the whole image, in our approach, the image is decomposed using FABEMD method into a series of BIMFs and residue, while the watermark takes place into one of BIMFs. The experiments indicate that our developed FABEMD-DCT-DWT technique can achieve high imperceptibility while showing good robustness. Furthermore, the proposed method is compared with the DCT-DWT based method, and the experiment results confirm that the proposed method shows better performances.

The authors of the present paper previously introduced a fast approximation framework for the graph edit distance problem. The basic idea of this approximation is to build a square cost matrix

C

 = (

c

ij

), where each entry

c

ij

reflects the cost of a node substitution, deletion or insertion plus the matching cost arising from the local edge structure. Based on

C

an optimal assignment of the nodes and their local structure is established in polynomial time. Yet, this procedure considers the graph structure only in a local way, and thus, an overestimation of the true graph edit distance has to be accepted. The present paper aims at reducing this overestimation by means of an additional greedy search strategy that builds upon the initial assignment. In an experimental evaluation on three real world data sets we empirically verify a substantial gain of distance accuracy while run time is nearly not affected.

A new framework to cluster images based on Genetic Algorithms (GAs) is proposed. The image database is represented as a weighted graph where nodes correspond to images and an edge between two images exists if they are sufficiently similar. The edge weight expresses the level of similarity of the feature vectors, describing color and texture content, associated with images. The image graph is then clustered by applying a genetic algorithm that divides it in groups of nodes connected by many edges with high weight, by employing as fitness function the concept of weighted modularity. Results on a well-known image database show that the genetic approach is able to find a partitioning in groups of effectively similar images.

The paper presents the algorithm of an automatic evaluation of the fluorescent in situ hybridization (FISH) images in order to determine HER2 status of the breast cancer samples. The algorithm is based on the accurate measurement of the red/green spot ratio (the ratio of HER2/CEN17) per cell nucleus. The main points of this algorithm is an accurate detection of the nuclei of cells and application of the color map to detect the red and green spots localized in these cells. The results of the numerical experiments concerning these two problems, as well as the assessment of the positivity or negativity of the considered cases are presented and discussed in the paper. They confirm the high efficiency of the proposed solution.

Denoising of images perturbed by non-standard noise models (e.g., Poisson or Gamma noise) can be often realized by a sequence of penalized weighted least-squares minimization problems. In the recent past, a variety of first-order algorithms have been proposed for convex problems but their efficiency is usually tested with the classical least-squares data fidelity term. Thus, in this manuscript, first-order state-of-the-art computational schemes are applied on a total variation penalized weighted least-squares denoising problem and their performance is evaluated on numerical examples simulating a Poisson noise perturbation.

One of the most important challenges of dealing with digital images acquired under uncontrolled conditions is the capability to assess if the image has enough quality to be further analyzed. In this scenario, blur can be considered as one of the most common causes for quality degradation in digital pictures, particularly in images acquired using mobile devices. In this study, we collected a set of 78 features related with blur detection and further analyzed its individual discriminatory ability for two dermatologic image datasets. For the dataset of dermoscopic images with artificially induced blur, high separation levels were obtained for the features calculated using DCT/DFT and Lapacian groups, while for the dataset of mobile acquired images, the best results were obtained for features that used Laplacian and Gradient groups.

Interpolations are among the most important tools for image processing. However, whether they are used for image compression and reconstruction purposes or for the increase of the image resolution along vertical, horizontal or both dimensions, the induced interpolation errors are often only qualitatively and a posteriori described. In this paper, we propose to extend a method used in an OFDM context to achieve a quantitative a priori estimation of interpolation errors. As shown by simulations, this estimation proves to be consistent with a posteriori error and quality measurements, such as mean square error (MSE) and peak signal-to-noise ratio (PSNR).

Usual cameras can gather only a small interval of intensities found in high dynamic range scenes. This fact leads to loss of details in acquired images and apparition of under or overexposed pixels. A popular approach to deal with this problem is to take several images differently exposed and fuse them into one single image. The exposure fusion is mostly performed as a weighted average between corresponding pixels. Weighting all pixels of the exposure bracketing slows the fusion process and makes realtime acquisition difficult. In this paper we present a fast exposure fusion method based on histograms segmentation. The segmentation phase reduces considerably the computations while preserving competitive fusion quality. We present also an automatic way to take enhanced exposures for fusion, using the segmented regions. Subjective and objective comparisons are conducted to prove the effectiveness of our method.

In this paper, we provide a new non-local method for image denoising. The key idea we develop is to denoise the components of the image in a well-chosen moving frame instead of the image itself. We prove the relevance of our approach by showing that the PSNR of a grayscale noisy image is lower than the PSNR of its components. Experiments show that applying the Non Local Means algorithm of Buades et al. [5] on the components provides better results than applying it directly on the image.

Optimal spectral bands selection is a primordial step in multispectral images based systems for face recognition. In this context, we select the best spectral bands using a multilinear sparse decomposition based approach. Multispectral images of 35 subjects presenting 25 different lengths from 480nm to 720nm and three lighting conditions: fluorescent, Halogen and Sun light are groupped in a 3-mode face tensor T of size 35×25×2 . T is then decomposed using 3-mode SVD where three mode matrices for subjects, spectral bands and illuminations are sparsely determined. The 25×25 spectral bands mode matrix defines a sparse vector for each spectral band. Spectral bands having the sparse vectors with the lowest variation with illumination are selected as the best spectral bands. Experiments on two state-of-the-art algorithms, MBLBP and HGPP, showed the effectiveness of our approach for best spectral bands selection.

Tone mapping operators (TMOs) convert high dynamic range (HDR) images to low dynamic range (LDR) images and are important because of the limitations of many standard display devices. Even though the quality of the resulting LDR image mostly depends on TMO parameter values, in this paper it is shown that it can be further improved by using alternative definitions of the luminance channel, which TMOs process. A new model of the luminance channel calculation that increases the resulting LDR image quality is also proposed. The main advantage of the new model is that the TMOs that produce results of lower quality can be made to produce results of significantly higher quality.

In this paper a novel tone mapping operator (TMO) based on the Light Random Sprays Retinex (LRSR) algorithm is presented. TMOs convert high dynamic range (HDR) images to low dynamic range (LDR) images, which is often needed because of the display limitations of many devices. The proposed operator is a local operator, which retains the qualities of the LRSR and overcomes some of its weaknesses. The results of the execution speed and quality tests are presented and discussed and it is shown that on most of the test images the proposed operator is faster and in terms of quality as good as Durand’s TMO, one of the currently best TMOs. The C++ source code of the proposed operator is available at

http://www.fer.unizg.hr/ipg/resources/color_constancy/

.

This paper presents an adaptation of the shock filter on weighted graphs using the formalism of Partial difference Equations. This adaptation leads to a new morphological operators that alternate the nonlocal dilation and nonlocal erosion type filter on graphs. Furthermore, this adaptation extends the shock filters applications to any data that can be represented by graphs. This paper also presents examples that illustrate our proposed approach.

Clothing, carrying conditions, and other intra-class variations, also referred as ”covariates”, affect the performance of gait recognition systems. This paper proposes a supervised feature extraction method which is able to select relevant features for human recognition to mitigates the impact of covariates and hence improve the recognition performance. The proposed method is evaluated using CASIA Gait Database (Dataset B) and the experimental results suggest that our method yields attractive results.

We propose an efficient solution for preserving the motion boundaries in variational optical flow methods. This is a key problem of recent TV-

L

1

methods, which typically create rounded effects at flow edges. A simple strategy to overcome this problem consists in inhibiting the smoothing at high image gradients. However, depending on the strength of the mitigating function, this solution may derive in an

ill-posed

formulation. Therefore, this type of approaches is prone to produce instabilities in the estimation of the flow fields. In this work, we modify this strategy to avoid this inconvenience. Then, we show that it provides very good results with the advantage that it yields an unconditionally stable scheme. In the experimental results, we present a detailed study and comparison between the different alternatives.

Methods for classification of images are of an important research area in image processing and pattern recognition. In particular, image classification is playing an increasingly important role in medicine and biology with respect to medical diagnoses and drug discovery, respectively. This paper presents a new method for image classification based on the frameworks of fuzzy sets and geostatistics. The proposed method was applied to the automated detection of regions of mitochondria in microscope images. The high correction rate of detecting the locations of the mitochondria in a complex environment obtained from the proposed method suggests its effectiveness and its better performance than several other existing algorithms.

In this paper, we propose a new set of Charlier-Tchebichef invariant moments. This set is derived algebraically from the geometric invariant moments. The presented approach is tested in several well known computer vision datasets including moment’s invariability and classification of objects. The performance of these invariant moments used as pattern features for a pattern classification is compared with Tchebichef-Krawtchouk, Tchebichef-Hahn and Krawtchouk-Hahn invariant moments.

We present a real-time multi-target tracking system that effectively deals with false positive detections. In order to achieve this, we build a novel motion model that treats false positives on background objects and false positives on foreground objects such as shoulders or bags separately. In addition we train a new head detector based on the Aggregated Channel Features (ACF) detector and propose a schema that includes the identification of true positives with the data association instead of using the internal decision-making process of the detector. Through several experiments, we show that our system is superior to previous work.

Optical flow is an ill-posed underconstrained inverse problem. Many recent approaches use total variation (TV) to constrain the flow solution to satisfy color constancy. We find that learning a 2D overcomplete dictionary from the total variation result and then enforcing a sparse constraint on the flow improves the result. A new technique using partially-overlapping patches accelerates the calculation. This approach is implemented in a coarse-to-fine strategy. Our results show that combining total variation and a sparse constraint from a learned dictionary is more effective than total variation alone.

Facade labeling, namely semantically segmenting the facade images, requires exploiting the facade regularity. To model the regularity, this paper proposes a novel matrix multiplication based formulation. In the model, the regularity is described as generalized translation symmetry (GTS) which enables varying distances between the repeated elements. Moreover, an explicit and intuitive formulation via matrix multiplication is also derived for the GTS. That is, the symmetry is interpreted as the product of a repetitive pattern and two block matrices. These two block matrices respectively represent the vertical and horizontal repetitions. Based on the formulation, facade labeling is reformulated into factorizing the facade to calculate the block matrices. An alternating optimization algorithm is thus developed to solve the matrix factorization problem, where dynamic programming is used to optimize the block matrices. Extensive experiments demonstrate the fidelity of our model and the efficiency of the algorithm.

In this paper we address the problem of matching video clips, each of which contains an instance from the same entity but undergoing transformation. To this end we formulate the problem as manifold matching by measuring the similarity between multiple manifolds, each represents a video clip. This work is novel in that it does not require a template or training. Instead it analyses the video by characterising the spatio-temporal information embedded in a frame sequence. Firstly the spatial Isomap is extended to spatio-temporal graph-based manifold embedding in order to discover the underlying structure of a video stream. Secondly linear models are extracted from each manifold through a hierarchical clustering method. The problem is then formulated as finding the distances between a pair of subspaces, each from one of the manifold. Experiment on Flicker dataset proved that the scheme was able to improve the search and retrieval performance over conventional approaches.

In this paper we present an approach for the segmentation and recognition of petroglyphs from images of rock art reliefs. To identify symbols we use a shape descriptor derived by 2-D Fourier transform, which is independent to scale and rotation, and robust to shape deformations. The efficacy of the algorithm has been validated by testing it with scenes and test images extracted from the archeological site located in Mount Bego (France). The results have been compared with those obtained by other descriptors.

Road detection is a significant task for the development of intelligent vehicles as well as advanced driver assistance systems (ADAS). For the past decade, many methods have been proposed. Among these approaches, one of them uses log-chromaticity space based illumination invariant grayscale image. However, errors in road detection could occur due to over saturation or under saturation, especially in weak lighting situations. In this paper, a new approach is proposed. It combines fisheye image information (in log-chromaticity space and in Lab color space) and laser range finder (LRF) measurements. Firstly, road is coarsely detected by a classifier based on the histogram of the illumination invariant grayscale image and a predefined road area. This fisheye image based coarse road detection is then faced to LRF measurements in order to detect eventual conflicts. Possible errors in coarse road detection can then be highlighted. Finally, in case of detected conflicts, a refined process based on Lab color space is carried out to rule out the errors. Experimental results based on real road scenes show the effectiveness of the proposed method.

Because of their visual characteristic which consists of black background versus white foreground, extracting relevant descriptors from medical X-ray images remains a challenging problem for medical imaging researchers. In this paper, we conduct an empirical comparison of several feature descriptors in order to evaluate their efficiency in content based X-ray image retrieval. We use a collection of X-ray images from ImageCLEF2009 data set in order to assess the performance of nine different visual descriptors with respect to different X-ray image categories.

The basic idea behind the energy-transfer features (ETF) is that the appearance of objects can be successfully described using the function of energy distribution in the image. This function has to be reduced into a reasonable number of values. These values are then considered as the vector that is used as an input for the SVM classifier. The process of reducing can be simply solved by sampling; the input image is divided into the regular cells and inside each cell, the mean of the values is calculated. In this paper, we propose an improvement of this process; the Discrete Cosine Transform (DCT) coefficients are calculated inside the cells (instead of the mean values) to construct the feature vector. In addition, the DCT coefficients are reduced using the Principal Component Analysis (PCA) to create the feature vector with a relatively small dimensionally. The results show that using this approach, the objects can be efficiently encoded with the relatively small set of numbers with promising results that outperform the results of state-of-the-art detectors.

A new approach for shape comparison based on skeleton matching is proposed. The skeleton of a binary image is encoded as a series of primitives (chain of primitives). Traditionally, a primitive is a pair of numbers, the first one is the length of the some edge and the second one is the angle between this and the neighbour edges. As a novelty we offer to calculate the Legendre polynomial coefficients to describe the width of shape and incorporate them as the third vector component into the primitive. The procedure of the alignment of two primitive chains is suggested and the pair-wise comparison function based on optimal alignment is built. Experiments with developed comparison function on the real-world dataset of medicinal leaves show that the results of classification are appropriate considering the difficulty of the task and disadvantages of the database.

Most of the computational color constancy approaches are based on the assumption of a uniform illumination in the scene which is not the case in many real world scenarios. A crucial ingredient in developing color constancy algorithms which can handle these scenarios is a dataset of such images with accurate illumination ground truth to be used both for estimating the parameters and for evaluating the performance. Such datasets are rare due to the complexity of the procedure involved in capturing them. To this end, we provide a framework for capturing such dataset and propose our multi-illuminant scene image dataset with pixel-wise accurate ground truth. Our dataset consists of 6 different scenes under 5 illumination conditions provided by two or three distinctly colored illuminants. The scenes are made up of complex colored objects presenting diffuse and specular reflections. We present quantitative evaluation of the accuracy of our proposed ground truth and show that the effect of ambient light is negligible.

In this paper we assess three standard approaches to build irregular pyramid partitions for image retrieval in the

bag-of-bags of words model

that we recently proposed. These three approaches are: kernel

k

-means to optimize multilevel weighted graph cuts, Normalized Cuts and Graph Cuts, respectively. The

bag-of-bags of words

(

BBoW

) model, is an approach based on irregular pyramid partitions over the image. An image is first represented as a connected graph of local features on a regular grid of pixels. Irregular partitions (subgraphs) of the image are further built by using graph partitioning methods. Each subgraph in the partition is then represented by its own signature. The

BBoW

model with the aid of graph, extends the classical bag-of-words (

BoW

) model, by embedding color homogeneity and limited spatial information through irregular partitions of an image. Compared to existing methods for image retrieval, such as Spatial Pyramid Matching (

SPM

), the

BBoW

model does not assume that similar parts of a scene always appear at the same location in images of the same category. The extension of the proposed model to pyramid gives rise to a method we name

irregular pyramid matching

(

IPM

). The experiments on

Caltech-101

benchmark demonstrate that applying kernel

k

-means to graph clustering process produces better retrieval results, as compared with other graph partitioning methods such as Graph Cuts and Normalized Cuts for

BBoW

. Moreover, this proposed method achieves comparable results and outperforms

SPM

in 19 object categories on the whole

Caltech-101

dataset.

Face recognition performance achieves high accuracy in close proximity. However, great challenges still exist in recognizing human face at long distance. In fact, the rapidly increasing need for long range surveillance requires a passage from close-up distances to long distances which affects strongly the human face image quality and causes degradation in recognition accuracy. To address this problem, we propose in this paper, a multispectral pixel level fusion approach to improve the performance of automatic face recognition at long distance. The main objective of the proposed approach is to formulate a method to enhance the face image quality as well as the face recognition rate. First, visible and near-infrared images are decomposed into a different bands using discrete wavelet transform. Then, the fusion process is performed through the singular value decomposition and principal component analysis. The results highlight further the still challenging problem of face recognition at long distance, as well as the effectiveness of our proposed approach as an alternative solution to this problem.

Segmentation of vessels in retinal images has become challenging due to the presence of non-homogeneous illumination across retinal images. This paper develops a novel adaptive thresholding technique based on local homogeneity information for Retinal vessel segmentation. Different types of local homogeneity information were investigated. An experimental evaluation on DRIVE database demonstrates the high performance of all types of homogeneity considered. An average accuracy of 0.9469 and average sensitivity of 0.7477 were achieved. While compared with widely previously used techniques on DRIVE database, the proposed adaptive thresholding technique is superior, with a higher average sensitivity and average accuracy rates in the same range of very good specificity.

An object of interest (OOI) in an image usually consists of visually coherent regions that, together, encompass the entire OOI. We use Multiple Instance Learning (MIL) to determine which regions in an over-segmented image are part of the OOI. In the learning stage, a set of over-segmented images containing, i.e., positive, and not containing, i.e., negative, an instance of the OOI is used as training data. The resulting learned prototypes represent the visual appearances of OOI regions. In the OOI segmentation stage, the new image is over-segmented and regions that match prototypes are merged. Our MIL method does not require prior knowledge about the number of regions in the OOI. We show that, with the coexistence of multiple prototypes corresponding to the regions of the OOI, the maxima of the formulation are good estimates of such regions. We present initial results over a set of images with a controlled, relatively simple OOI.

Poisson surface reconstruction is applied as an efficient technique to create a watertight surface from oriented point samples acquired with 3D range scanner or dense multi-view stereopsis. With non-homogenously distributed noisy sparse point cloud, Poisson surface reconstruction suffers from the problems of either over-smoothness, or large area of unrecognizable reconstructed surface. We present a novel three-step framework to provide a 3D mesh which better approximates the real surface of the object based on an iterative energy minimization process. The experimental results show the feasibility of the proposed approach on real image datasets.

Random Indexing is an incremental indexing approach that simultaneously performs an implicit Dimensionality Reduction and discovers higher order relations among features lying in the vector space. The present work explores the possible application of Random Indexing in discovering feature contexts from image data, based on their semantics. We propose an automatic approach of image parsing, feature extraction, indexing and clustering, showing that the Feature Space model based on Random Indexing captures the semantic relation between similar features through a mathematical model. Experiments show that the proposed method achieves good clustering results on the large Corel database of 599 different semantic concepts.

Geometric Constraint Solving Problems (GCSP) are nowadays routinely investigated in geometric modeling. The 3D Pentahedron problem is a GCSP defined by the lengths of its edges and the planarity of its quadrilateral faces, yielding to an under-constrained system of twelve equations in eighteen unknowns. In this work, we focus on solving the 3D Pentahedron problem in a more robust and efficient way, through a new formulation that reduces the underlying algebraic formulation to a well-constrained system of three equations in three unknowns, and avoids at the same time the use of placement rules that resolve the under-constrained original formulation. We show that geometric constraints can be specified in many ways and that some formulations are much better than others, because they are much smaller and they avoid spurious degenerate solutions. Several experimentations showing a considerable performance enhancement (×42) are reported in this paper to consolidate our theoretical findings.

Surface reconstruction towards 3D data is a popular case study in the field of computer graphics. Although many methods are able to solve surface reconstruction problems, but limitations are still appeared. The limitations of Kohonen Self Organizing Map (KSOM) model in closed surface was handled by introducing Cube KSOM (CKSOM) model. However, the CKSOM model output is not in industrial standard format because NURBS are mostly used as surface representation in computer aided geometric design. Furthermore, NURBS surface approximation result will be affected by the parameterization methods. Therefore, the aims are to test and apply NURBS surface approximation on the CKSOM model output and to obtain less surface errors using different parameterization methods. Based on the result, NURBS was proven to be able to apply on the CKSOM model output and uniform parameterization method was proven to be the best method compared to others based on the surfaces error obtained.

3D modeling and mesh editing for objects from orthographic projections of engineering drawing has been very common and convenient. People has been accustomed to using such simple and fast sketching way to express their visualized thought. In this paper we present a method for model designers by sketching on the orthographic projection to control the deformation of surface meshes. Our system lets the user easily choose the region of interest for constraining the deformation area, and then sketch the desired contour. For a given new sketch on the orthographic projection,the system automatically deform the 3D model by using Laplacian deformation, which makes the contours of 3D model fit the contour line of the users sketching. Various examples have validated the effectiveness of our proposed method, which can be regarded as a effective method in 3D model editing.

In computer aided celiac disease diagnosis, endoscopes with wide-angle lenses are deployed which induce significant lens distortions. This work investigates an approach to automatize the estimation of the lens distortion, without a previous camera calibration. Knowing the discriminative power of all sensible distortion configurations, the model parameters are estimated. As the achieved parameters are not highly precise, moreover, we investigate the effect of approximative distortion correction on the classification accuracy. Particularly, we identify one simple but especially for certain features highly effective approximative distortion model.

Recent advancement in biomedical image processing using Magnetic Resonance Imaging (MRI) makes it possible to detect and localize brain tumors with ease. However, reliable classification of brain tumor types using MRI still remains a challenging problem. In this paper we propose a sparse representation based approach to successfully classify tumors in brain MRI. We aim to classify brain scans into eight (8) different categories with seven (7) indicating different tumor types and one for normal brain. This allows the proposed approach to not only classify brain tumors but also to detect their existence. The proposed classification approach is validated using Leave 2-Out cross-validation technique. The result obtained from the proposed approach is then compared with a recent technique presented in literature. The comparison clearly shows that the proposed approach outperforms the existing technique both in terms of accuracy and number of classes being employed.

Since the early 1990s, speaker adaptation have become one of the intensive areas in speech recognition. State-of-the-art batch-mode adaptation algorithms assume that speech of particular speaker contains enough information about the user’s voice. In this article we propose to allow the user to manually verify if the adaptation is useful. Our procedure requires the speaker to pronounce syllables containing each vowel of particular language. The algorithm contains two steps looping through all syllables. At first, LPC analysis is performed for extracted vowel and the LPC coefficients are used to synthesize the new sound (with a fixed pitch period) and play it. If this synthesized sound is not perceived by the user as an original one then the syllable should be recorded again. At the second stage, speaker is asked to produce another syllable with the same vowel to automatically verify the stability of pronunciation. If two signals are closed (in terms of the Itakura-Saito divergence) then the sounds are marked as ”good” for adaptation. Otherwise both steps are repeated. In the experiment we examine a problem of vowel recognition for Russian language in our voice control system which fuses two classifiers: the CMU Sphinx with speaker-independent acoustic model and Euclidean comparison of MFCC features of model vowel and input signal frames. Our results support the statement that the proposed approach provides better accuracy and reliability in comparison with traditional MAP/MLLR techniques implemented in the CMU Sphinx.

Agricultural products are vital to the survival of man and in its larger scale of production could contribute to the economy of a nation. Farmers who are at the fore front of Agriculture are faced with several challenges. This paper identifies some of the challenges and in particular focuses on leaf diseases that affect crops. A disease if not well addressed could destroy crops and hence a loss in investment. Using contemporary technology that is affordable and portable, such as mobile devices, this paper develops a framework through which intervention by an expert system on crop diseases could be implemented.

Speech encryption is becoming more and more essential as the increasing importance of multimedia applications and mobile telecommunications. However, multimedia encryption and decryption are often computationally demanding and unpractical for power-constrained devices and narrow bandwidth environments. In this paper an encryption scheme for AM-WB ITU-T G. 722.2 speech based Arnold cat Map is presented analyzed and evaluated using objective and subjective tests for the 8 modes of the AMR-WB ITU-T G.722.2. Simulation results show that AMR-WB ITU-T G.722.2 based Arnold cat Map encryption is very efficient since the encrypted speech is similar to a white noise. The perceptual evaluation of speech quality (PESQ) and enhanced modified bark spectral distortion (EMBSD) tests for speech speech extracted from TIMIT database confirm the efficiency of the presented scheme.

Several research studies have shown that the robustness and performance of speech recognition systems can be improved using physiologically inspired filterbank based on Gabor filters. In this paper, we proposed a feature extraction method based on 59 two-dimensional Gabor filterbank. The use of these set of filters aims to extracting specific modulation frequencies and limiting the redundancy on feature level. The recognition performance of our feature extraction method is evaluated in isolated words extracted from TIMIT corpus. The obtained results demonstrate that the proposed extraction method gives better recognition rates to those obtained using the classic methods MFCC, PLP and LPC.