Hyperspectral imaging provides spectral information for every pixel in an image, though the data is less reliable for low levels of illuminance. Merging hyperspectral imaging with High Dynamic Range (HDR) imaging can improve the loss of accuracy in low-lit environments by compensating for dark pixels with high exposure values. A study involving a light-booth with a maximum of 135 cd/m(2) shows a significant improvement in spectral and photometric accuracy of the spectral data using HDR techniques. The sRGB image quality also improves significantly. The HDR merging method using a mean or median value of valid pixels also significantly improves results as compared to the first correct pixel. The results from this study provide a promising outlook for the use of hyperspectral imaging in poorly lit scenes and should be further investigated.
In recent studies, alongside introducing new approaches for color constancy, we have also focused on improving existing techniques and introducing new perspectives. Our motivation is the idea that investigating different strategies, concepts, and their combinations that have not been analyzed in this field in detail yet, might help us to find simple, effective, and cost-efficient solutions. Thereupon, we utilized observations we obtained from our algorithms to analyze how we can enhance the performance of well-known learning-free methods. We demonstrated why using salient pixels, performing block-based operations, and carrying out scale-space computations benefit color constancy approaches significantly and provide a notable performance increase. In this study, we make use of our recent observations on learning-free algorithms to analyze if they are also beneficial for enhancing the performance of a learning-based color constancy model. According to our evaluation, all of our observations contribute to the performance of a convolutional neural network model and increase its effectiveness in estimating the illuminant. Thus, the contribution of these strategies in learning-based models should be further investigated to improve their performance with simple yet effective solutions.
In device-dependent color spaces, the Euclidean distance often fails to accurately represent perceived color differences, making these color spaces unsuitable for image similarity measurements. Conversely, certain device-independent color spaces, known as uniform color spaces (UCS), do align the Euclidean distance with perceived color differences. Among these, CAM16-UCS is currently considered the leading model. However, applications such as gamut mapping require additional properties, such as the preservation of planes of constant hue, which CAM16-UCS lacks. We present PCS23-UCS, a novel UCS that addresses this limitation. PCS23-UCS is specifically designed to preserve the central bundle of planes in the color space while maintaining state-of-the-art uniformity for small color differences.
Intrinsic image decomposition enables us to estimate the low-level features of images. Due to the benefits it provides and the challenges it holds, intrinsic image decomposition has been extensively studied over more than five decades. It can be utilized in various computer vision and computer graphics pipelines to improve the efficiency of tasks such as object classification and recoloring, and image segmentation. In this study, we introduce an algorithm for reflectance and shading estimation, offering a simple yet effective solution to the ill-posed intrinsic image decomposition problem. Our learning-free method leverages a combination of the fundamentals of the Retinex theory, scale-space computations, and superpixel segmentation. The assumptions of the Retinex theory enable us to provide a straightforward solution to a complex problem, while scale-space computations allow us to highlight low-level features and superpixel segmentation helps us to preserve local information. We evaluated our algorithm that mainly focuses on single objects on three benchmarks, namely, MIT Intrinsic Images, Bonneel, and MPI Sintel, which either consist of single objects or complex scenes having different characteristics. According to our experiments, our algorithm provides competitive results compared to the other methods.
Colorimetry can be computed for a given observer, illuminant and a surface, using the standard CIE formulae, resulting in a discrete point in three dimensional space, the CIE XYZ tristimulus value. While other representations are also possible (such as CIE LAB, LCh etc.), they all ultimately share one characteristic, that color is represented as a point in 3D space. However, there is known uncertainty and variation in every one of the stimuli, and quantities from which colorimetry is computed and perceptual attributes predicted. Likewise, there is uncertainty in perceptual evaluation – not only as a function of stimulus properties (size, context) but also fundamental variability between observers. The representation of color as a discrete point in 3D however obscures this underlying variability. In this paper an alternative is put forth, which claims that representing colors as points is both reductive and inaccurate when it comes to reflecting real-world performance and also – perhaps counterintuitively – introduces unnecessary complexity when designing optimization problems. Not only is it more correct to consider colorimetry in terms of ranges or probability distributions, it also leads to more realistic ways to talk about, e.g., whether or not two samples match, whether or not a colorimetry is in-gamut or not, and ultimately also helps in formulating color and machine learning models that open up new possibilities in using the latest AI techniques.
Hyperspectral imaging offers high spectral and spatial resolution, but its high costs and time-consuming nature make it difficult to use. Spectral Filter Array (SFA) imaging presents an alternative, offering high spectral resolution, user-friendliness, and affordability, but at the cost of limited spatial resolution. This paper presents an approach to address this trade-off, starting with raw overlapping frames from spectral videos, followed by a demosaicking network process before tackling the stitching problem. Our experiments on various spectral videos, supported by image quality metrics and qualitative demonstrations, indicate that this approach effectively enhances the spatial resolution of spectral images while reducing artifacts. The integration of the demosaicking and the stitching provides a robust solution for spectral video applications, paving the way for further advancements in panoramic spectral image stitching.
Color difference evaluation is an important topic in colorimetry, and it has led to many methods of calculations over the years, which still aim to improve accuracy. This study aims to explore the efficacy of existing color difference measures in evaluating perceived color differences under varying contextual conditions. A novel experiment was conducted wherein participants evaluated color pairs against different backgrounds. This contribution presents the preliminary results of this study which compare the perceived color on varying backgrounds with some of the most used color differences. Overall, the results show that the precision gained by improving pointwise color metrics is pointless due to the high variability of color perception in varying contexts caused by spatial mechanisms of the human visual system. In conclusion, this study highlights the importance of creating alternative methods to measure color differences that reflect how we see colors in real life. Such a challenge could improve color research and its applications across different fields.
Developing accurate methods for differentiating benign vs. malignant pulmonary nodules on CT is crucial for the correct management of patients referred for suspicious lung cancer. In this context deep learning by convolutional neural networks (CNN) has been gaining ground as an alternative to conventional methods based on feature engineering, although the use of CNN is often hampered by the lack of sufficiently large datasets for training. Herein we explore the effectiveness of deep features from pre-trained convolutional networks ‘off-the-shelf’ to discriminate benign vs. malignant lung nodules from CT images. To this end we compare three approaches (two classic, one novel) for pseudo-colour image generation which allow the grey-scale CT data to be fed into CNN models designed for and trained on colour images. The classification performance was estimated by internal and cross-validation using two independent datasets (LIDC-IDRI and LUNGx), giving four experimental conditions altogether. Conventional radiomics features were used as baseline reference. The best accuracy achieved by deep features in the four experimental conditions was respectively 88.5
Improving the quality of non uniformly lighted images is particularly hard due to the presence of regions that, having different brightness, requires different degrees of enhancement. The recently published algorithm REK proposes an interesting solution for enhancing images with abrupt changes of light intensity. REK linearly up-scales the image brightness to increase the quality of dark regions and combines the image with up-scaled brightness with the input one to preserve the quality of the bright regions. The up-scaling parameter α is estimated unsupervisedly based on the segmentation of brights and dark regions. This estimate has two main disadvantages: first, it makes REK dependent on the segmentation algorithm; second, the segmentation may be adversely affected by noise often present in badly illuminated images. To overcome these issues, this work proposes a new estimation of α based on the comparison of image Sobel gradient with an enhanced contrast, specifically the Milano Retinex contrast, which - as the name suggest - is inspired by the principles of Retinex theory. The new estimation of α provides good results, with reduced artifacts and saturation effects.
Forgery detection in fine art necessitates collaboration among art historians, conservators, scientists, and forensic experts. Traditional methods, which rely on expert visual assessment and scientific analysis, are effective but often time-consuming and costly. Recent advancements in technology have introduced image analysis and machine learning techniques, offering efficient and precise alternatives for detecting art forgeries. Excluding material and chemical-based clues, assessing the authenticity of paintings can be broadly modeled along three visual dimensions: color, brushstrokes, and contents. This paper examines the efficacy of using color as a feature for determining the authenticity of paintings. We utilize machine learning algorithms to analyze the color palettes of over 100,000 digital images from approximately 1,500 artists. We compactly represented paintings through their color palettes and different other color-based features. Our experiments, designed as verification tasks, explore the potential of color-based features to verify the authorship of the artworks
The advancements in Light-Emitting Diodes (LEDs) have allowed spectrally tunable light sources to gain attention in many fields of research thanks to their ability to produce a specific light output. However, LED outputs can fluctuate with temperature, and aging components can lead to noticeable discrepancies in light characteristics. This study thoroughly examines the Telelumen Dittosizer light player LED panel to exemplify a commercially available device and the associated challenges in predicting and stabilizing its output. Then, we introduce an innovative algorithm aimed at addressing such a stabilization challenge, based on a straightforward characterization procedure along with an external spectrometer. The accuracy of the algorithm was validated with different inputs, achieving a _E,2000 lower than 0.5. Our findings demonstrate the ability to stabilize the spectral power distribution for a minimum of 30 min. The proposed algorithm is hardware-independent and adaptable to any combination of spectrally tunable light sources and spectrometers.
This study investigates the robustness and precision of 2D human pose estimation techniques, particularly feature-based and 3D shape-based methods, against the backdrop of varying color illumination conditions simulated by chromatic adaptation, as well as varying spatio-temporal dynamics. While the AIST++ dance videos dataset serves as the primary data, the insights gained are pertinent to broader contexts like sports, highlighting the critical influence of lighting on the performance of pose estimation technologies.
Neural networks are now the standard solution to many computer vision problems. Their generalization ability enables them to successfully address various tasks in computational photography, such as enhancement, restoration, and color constancy. However, their performance is highly dependent on the illumination conditions of the training images. When faced with test images under different illuminant conditions, these networks often struggle to perform their tasks correctly. In this paper, we investigate the efficacy of illuminant equivariant neural networks for the illuminant estimation task, which is crucial for computational color constancy. These networks are equivariant to the photometric transformations that characterize changes in lighting conditions. They achieve this capability through mathematical derivation rather than specific augmentation during training. We implemented the equivariant versions of state-of-the-art neural networks for illuminant estimation and tested them on the NUS dataset. The results demonstrate that the equivariant networks maintain stable performance even with significant changes in illumination, whereas the original standard networks exhibit a serious degradation in their accuracy.
Nowadays deep networks provide excellent results in the context of object segmentation. Available models have been trained on common objects and are not designed to segment specific objects such as fruits or vegetables. In order to help breeders to accelerate and to modernize the process of agriculture products phenotyping, it is necessary to fine tune general models on specific species. Nevertheless, a minimum amount of annotations are required for this retraining step. In this paper, we propose a solution to minimize the annotation workload for each specie. The main idea consists in leveraging the annotations of one specie A in order to fine tune a model on a specie B with few annotations. For this purpose, we propose an Instance-based CycleGAN (ICG) that creates synthetic images of specie B along with corresponding annotations. By fine tuning a segmentation network with these synthetic images and annotations, we show that this network can obtain very good performance on the new specie B, without requiring to manually annotate a large amount of images for this specific specie B.
We propose spectral imaging methods for estimating fluorescence spectra emitted from plant grains and leaves. Two types of multiband imaging systems with six channels are constructed using ordinary off-the-shelf cameras; the first of these is for rice grains, wherein a mobile phone camera is rendered multiband to detect fluorescence emission in the visible wavelength region; the second one is for plant leaves, wherein a monochrome camera is used with additional optical filters to detect chlorophyll fluorescence in the red to far-red wavelength region. A statistical approach, inspired by the Wiener estimator is developed to estimate fluorescence emission from the image data. Here, the observations are modeled by accounting for noise in the imaging system, and the incident fluorescence spectra to the system are estimated from the observed image data. The noise variance is estimated on the basis of the fluorescence spectrum measurement. In the experiments, UV-light is used to illuminate real rice grains and green leaves; the emitted fluorescence spectra are estimated, and the reliability of the proposed methods is demonstrated.
In this article, I present two versions of a cellular automaton (CA) that evolve according to a set of rules derived from a well-known combinatorial structure: Pascal's Triangle. These CAs produce point sampling commonly observed in high-quality digital dithering and halftoning techniques. The first version is a probabilistic cellular automaton (PCA), which samples discrete values from Pascal's Triangle as a probability distribution. The second version is deterministic, where tone-dependent lattice paths are selected from Pascal's Triangle by analyzing the local points distribution in the halftone pattern. This deterministic version also offers significant computational gains over its probabilistic counterpart.
Optimization has been a thread running through the history of printing, from the motivation of Johannes Gutenberg developing moveable type to the present efforts to apply AI techniques to solving a broad range of printing challenges. The need for optimization here ranges from the solution of specific, well-defined challenges like the prediction of print color from as few measurements as possible, via the boosting of ink efficiency and printed image quality to the need for ease of use and automation necessitated by a desire to minimize human intervention in print manufacturing. While the former category of challenges is well suited to various flavors of machine learning, the latter becomes approachable thanks to the advent of large language models and other generative AI methods. A critical factor in this optimization spectrum is also the control domain in which printed output can be specified and therefore also adjusted. Here the HANS paradigm offers a clear benefit by making the space in which print is controlled linear, which allows for optimization processes performed on top of it to be directed at the specific challenges at hand, instead of also having to compensate for the ill-behavedness of conventional print control spaces. The paper concludes with a bit of blue-sky thinking about what the near future of print optimization may look like, where AI moves from particular, narrow application to providing an end-to-end infrastructure.
In colour science we are familiar with the Planckian locus, a curve in chromaticity space that is indexed by the temperature T of a blackbody radiator, that approximately delineates the colours of typical lights. In the CIE (u', v') chromaticity diagram the reddest (lowest-T) light intersects the Spectral locus at 830 nm; as T rises, the lights become progressively orange, yellow, white and blue with the bluest light being defined by an infinite colour temperature. Interestingly the bluest blue Planckian chromaticity is far from the spectral locus. The Wien locus is parameterised by a simpler equation than Planck and runs almost parallel with the Planckian locus. The loci are so close together that a temperature conversion brings corresponding chromaticities into a most complete coincidence. Though, the Wien locus is longer - extends more towards the short-wave part of the chromaticity diagram - than the Planck locus, for an infinite temperature. In this paper, we extend the Wien formula to allow negative temperatures. When we do this the Wien locus naturally extends all the way to intersect the spectral locus (at 360 nm). The extension to negative temperatures allows the Wien locus to model any reasonable light's chromaticity within its arc. We show that the extended Wien locus is continuous: negative and positive infinite colour temperatures converge to the same point. However, there is a substantial discontinuity at T = 0 evidenced by the large (u', v') difference between the blue and red ends of the Wien locus. A practical application of this result is discussed (via the theory of Locus Filters) to improve design of color-correction filters in photography.
This paper focuses on enhancing temporal color constancy in video sequences, ensuring that the result not only achieves color accuracy frame-by-frame but is also consistent over time. Our approach consists of a three-step process: per-frame illuminant estimation and correction, video stabilization to ensure temporal consistency, and consensus-based illuminant correction. By employing consensus-driven illuminant estimation over the result of temporal stabilization, we effectively mitigate spatial artifacts and concurrently enhance the overall stability of the sequence. Our method is tested using the Gray Ball and BCC datasets, showing the potential of integrating temporal stabilization with color correction processes to enhance the visual continuity of video content. While our primary objective is to reduce temporal flickering, a significant side effect of our approach is the improvement of color constancy accuracy across frames.
In this study, we focus on glossiness that changes when reproducing colorimetric images of real objects. We experimentally explore image reproduction that is perceptually equivalent to the real object in terms of glossiness, which is necessary to retain the glossiness that can change. In addition to direct comparison of perceptual glossiness of real objects and colorimetrically reproduced images, psychophysical experiments were conducted to directly compare edited images and real objects to reproduce perceptual glossiness. Because direct comparisons of real objects and colorimetric reproduction images showed that the perceptual gloss decreased due to imaging for most stimuli, psychophysical experiment was conducted using processed image to reproduce the perceptual glossiness. The results of both experiments were analyzed using GLCM features, a type of texture analysis, suggesting that contrast, dissimilarity, and entropy are more reproducible than colorimetric reproduced images, whereas homogeneity is less reproducible than colorimetric reproduced images. When homogeneity and ASM are smaller than the colorimetric reproduced image, the image tends to be reproduced with a glossy appearance more similar to the original image.
