ES2947034T3 - Method and device for evaluation of color similarity - Google Patents

Abstract

Method and device configured for the evaluation of color similarity between a first and second color, comprising the steps of: calculating one or more filters from said first and second color; determining that the first and second colors are different if at least one of the calculated filters is above a predetermined threshold for each filter; otherwise, calculating the CIEDE2000 color difference and determining that the first and second colors are different if the calculated CIEDE2000 color difference is above a predetermined threshold; otherwise, determining that the first and second colors are similar. A module filter can calculate the module of the difference vector between the first and the second color represented as 3D coordinates in the RGB color space, (Automatic translation with Google Translate, without legal value)

Description

DESCRIPTION

Method and device for evaluation of color similarity

technical field

The present disclosure refers to a method and a device for the evaluation of similarity of colors, specifically the comparison of two colors, applicable to areas, such as digital image processing or computer vision.

Background

Color differences are of great importance in many applications, especially when computer vision or imaging techniques are to be used. Examples of such application topics are mineral identification, dental ceramics, wine color analysis, corrosion detection and, in a more generic scope, image segmentation. The latter is widely used and is often a critical stage in image and video processing algorithms. Noise in the result of this step affects the results of the following processing, specifically in segmentation, detection or appearance matching, possibly preventing correct operation. Despite many years of research, image segmentation remains an unresolved research topic, in which color identification plays a key role. The CIEDE2000 [3] metric was recommended by CIE in 2000 for the industrial evaluation of color differences. However, its advantages have been diminished by the computational cost.

The CIEDE2000 [3] formulation is an international standard since 2014, which is incorporated herein as a reference in its entirety and in particular in its calculation of color differences, with reference to ISO CIE 11664-6:2014 (CIE S 014-6/E:2013), and has a strong focus on industrial applications. This metric has been receiving more and more attention, showing a high potential in the discrimination between colors, especially those with greater similarity. The use of the CIEDE2000 [3] (AE ⁰⁰ ) metric allows quantifying how humans perceive the difference between two colors. However, its widespread use has been hampered by the high computational cost, which prevents its application in many situations, especially those with real-time limitations. Although technological evolution has allowed greater computing capacity, there are a large number of situations that remain unfeasible due to high amounts of visual information or restricted hardware.

Color-based algorithms can be severely affected by lighting, and discrimination between closest colors is rarely easy. The CIEDE2000 metric is increasingly used in many applications and allows more or less precise discrimination depending on the selected metric threshold. The AE ⁰⁰ color difference formula was designed to characterize the perceptual color difference exhibited by humans and can be summarized in equations ¹ to 28:

where Kl, Kc and Kh are correction factors related to lightness, chroma and hue, respectively. In experiments, these values were set to 1, as defined in the CIE standard. The parameters Sl, Sc and Sh are weighting factors; AL', AC' and AH' are the differences in lightness, chroma and hue. Li, ai, bi and L ² , a ² , b ² are two colors in the CIELAB color space, with ISO 11664-4:2008 (CIE S 014-4/E:2007).

The edge detection methodology presented in [4] combined edge maps from different sources in order to extend depth information with limited spatial resolution. Some alternatives were shown and the one that provided the best results was the one comprised of a combination of edge maps for color difference and intensity. The color difference edge map was obtained by using the CIEDE2000 color difference, in order to obtain a perceptual color edge detection that follows the characteristics of the human visual system. Although this solution provided the best results, it came at the cost of high computational complexity. For color similarity decisions using AE ⁰⁰ , values equal to or less than 3 are typically applied; for values greater than 10, two colors are considered to be very different.

The CIEDE2000 metric [3] presents other concurrent metrics, specifically other metrics proposed by CIE such as CIE76 (also known as DeItaE CIELAB), CIE94 and DeItaE CMC. However, several studies have concluded that more robust results can be obtained using AE ⁰⁰ even with the default settings. Better results can be obtained by adjusting the factors in the formula to the characteristics of a given application.

The paper “Measuring Images: Differences, Quality and Appearance”, GARRETT M JOHNSON, DISSERTATION, 20030301, page(s):1 - 166, XP002444668, discloses a modular framework for creating a color image difference metric using ensembles of psychophysical data. This framework is based on the CIE color difference equations, and the S-CIELAB spatial extent with respect to the CIELAB color space.

The paper “Distance measures for color image retrieval”, Androutsos D, Plataniotiss K N, Venetsanopoulos A N, image processing, 1998, ICIP 98. Proceedings, Chicago International Conference 1998, IL. USA Oct 4-7 from 1998, 19981004; 19981004 - 19981007 Los Alamitos, CA, USA, IEEE Comput. Soc, US, discloses color-based retrieval of image databases using various vector distance metrics. This is based on color segmentation where only a few representative color vectors are extracted from each image and used as image indices. These vectors are then used with vector distance measurements to determine the similarity between a query color and a database image.

Document US2005248783A1 discloses a color characterization process that uses nonlinear regression analysis to characterize a color input device. The color input device is used to generate a bitmap of device-dependent values from a color target. The device-dependent color values bitmap is used to generate a forward model that maps device-dependent color values to color values in a device-independent color space using nonlinear regression analysis that minimizes a color difference metric between the reference color values and the set of device-dependent color values mapped through the advance model. The color difference metric is chosen to represent human-perceived color differences in device-independent color space. The performance of the nonlinear regression analysis can be improved by initializing the nonlinear regression analysis using a seed run model generated from a linear regression analysis.

These facts are disclosed for the purpose of illustrating the technical problem addressed by the present disclosure.

References

[1] Griffin, Gregory and Holub, Alex and Perona, Pietro. Caltech-256 object category dataset. 2007.

[2] Liaw, Andy and Wiener, Matthew. Classification and regression by randomForest. Rnews, 2(3):18-22, 2002.

[3] Luo, M Ronnier and Cui, Guihua and Rigg, B. The development of the CIE 2000 color-difference formula: CIEDE2000. Color Research & Application, 26(5):340-350, 2001.

[4] Schwarz, Sebastian and Sjostrom, Marten and Olsson, Roger. Improved edge detection for EWOC depth upscaling. 19th International Conference on Image Systems, Signals and Processing 2012 ( IWSSIP), pages 1-4, 2012. IEEE.

[5] Color names supported by all browsers. http://www.w3schools.com/colors/colors_names.asp. [Online; consulted on 04-08-2016].

[ ⁶ ] Dang-Nguyen, Duc-Tien and Pasquini, Cecilia and Conotter, Valentina and Boato, Giulia. RAISE: a raw images dataset for digital image forensics. Proceedings of the 6th ACM Multimedia Systems Conference, pages 219-224, 2015. ACM.

[7] Pecho, Oscar E and Ghinea, Razvan and Alessandretti, Rodrigo and Perez, Mara M and Della Bona, Alvaro. Visual and instrumental shade matching using CIELAB and CIEDE2000 color difference formulas. Dental Materials, 32(1):82-92, 2016.

General description

The invention is defined by a computer-implemented method for evaluating color similarity, according to claim ¹ , by a storage medium that includes program instructions to implement a color similarity evaluation device, according to claim 9, and by a color similarity evaluation device comprising a data processor, according to claim ¹⁰ .

The important advantages of the AE ⁰⁰ metric are often undermined by its complexity. When analyzing the AE ⁰⁰ formula (previously presented and described in detail in [3]), the high complexity of several of the terms is appreciated, which gives rise to a formula with a large computational cost. Therefore, the use of this technique in many application situations is usually unfeasible, especially when computational resources are more limited. In addition, applications that need to perform multiple AE ⁰⁰ calculations will be seriously affected by high computation times. As an example, two images with resolutions of 720 * 480 and 1920 * 1080 pixels are considered; in the first it will be necessary to process 0.3 * 10 ⁶ pixels and in the second 2.1 * 106, which means many more calculations. The computational cost drawback of AE ⁰⁰ becomes even more apparent when considering video, with multiple frames per second, and the increasingly pervasive 4K and ^8K image resolutions. To overcome this problem and extend the applicability of this formula and its benefits to more situations, a method and device for a significantly faster decision on color similarity is disclosed, while retaining the benefits and capabilities of AE ⁰⁰ .

In order to decrease computation times, the disclosed method and device apply a set of filter modules where simpler calculations and evaluation are performed: ( ¹ ) the two input colors are different and no more is needed. calculations; ( ² ) A decision cannot be made with certainty and further analysis is required. Typically, when comparing a pair of colors, these will be an input for the modules and AE ⁰⁰ will only be fully computed for pairs that may have higher perceptual similarity and have not had a decision made by the lower weight modules. . Thus, the number of times the full AE ⁰⁰ metric is computed, and consequently the total number of operations required, is effectively reduced, resulting in shorter processing times, while maintaining the same precision. This is supported by the fact that the expected result is not a precise distance measurement, but a decision about whether or not a pair of colors is similar. However, it is still necessary to identify and establish relationships between the filter modules and AE ⁰⁰ so that the results are equal to those obtained by applying only AE ⁰⁰ , thus maintaining the metric advantages.

Colors are usually represented in the RGB color space. This is especially true when it comes to images, which are typically acquired in the RGB color space. Since the AE ⁰⁰ metric uses the color space CIELAB often requires an additional color conversion step, which leads to additional calculations. It is advantageous for filter modules to avoid as far as possible color space conversions, more specifically, those that operate on the RGB color space. This space can be represented as a cube; thus the difference between two colors can be seen as a difference vector within space. With this idea, the disclosed method and device propose three properties to be explored: the magnitude of the vector differentiates between two colors; the angle between two color vectors; the dissimilarity of luminosity of the colors. Unlike the first two, the latter is not computed in RGB space, but requires color conversion (for example, converting to CIELAB luminosity using CIElA b -CIEXYZ conversion and ITU-R standard system values BT.709-6, as detailed below). In the rest of this document, these properties will be referred to as modulus, angle, and luminosity, respectively.

Given two colors in RGB, Ci = (n, gi, bi) and C ² = (r2, g2, b2), the angle, modulus, and luminosity properties are calculated as follows:

AL is given by the following equations,

where Ci and C ² are the normalized color vectors.

The calculations of equations 32 and 33 are carried out in each component of the vectors and are presented in this case in a vectorized form. The referenced properties have well-defined limits: the angle between two color vectors (in RGB color space) is limited between 0 and 90 degrees; the difference vector is bounded by 0 as the lower bound, if the vectors are equal, and 441.673 if they are at opposite vertices of the cube; the luminosity difference is limited by 0 and 100. The use of these properties is justified by: (1) their simplicity, especially when compared with AE ⁰⁰ as can be seen in equations 1 to 28; (2) its well-defined boundaries.

If a certain maximum threshold of AE ⁰⁰ is set, that is, if we define a value of AE ⁰⁰ above which two colors are considered to be perceptually different, then it is possible to reduce the number of calculations performed for the colors that are expected to be different. give a color difference greater than the defined limit, by thresholding the angle, modulus and brightness. Therefore, it is necessary to identify the relationships of angle, module and luminosity and the different results of the metric AE ⁰⁰ . To understand these relationships and the variability of the many parameters of AE ⁰⁰ , a set of approximately 5 million colors was used and the differences were calculated using the image data present in the Caltech 256 dataset as base pixels [1]. .

Figures 1, 2 and 3 show the relationship of AE ⁰⁰ with module, angle and luminosity, respectively. For each color pair, different thresholds of AE ⁰⁰ were taken into consideration, namely 3, 5 and 10. Color differences above the thresholds are represented in dark gray and light gray otherwise. Furthermore, the corresponding thresholds for module, angle and brightness are represented by a horizontal line. All pairs whose AE ⁰⁰ value is below the specified threshold are guaranteed to be also below the modulus, angle, and luminosity thresholds. As can be seen, for values less than the horizontal line a safe decision cannot be made based on the filter modules and additional calculations are required. Figures 1, 2 and 3 show that for modulus, angle and luminosity values above the horizontal line a precise decision can be made about their dissimilarity, that is, they can be classified as different using only the proposed filtering properties. In addition, the plots show that increasing the decision value defined for AE ⁰⁰ requires higher values for the modulus, angle, and luminosity thresholds. However, the number of color differences that can be filtered out is always very significant. Note that AE ⁰⁰ values greater than 7 are considered to represent clear color differences.

A method for evaluating color similarity according to claim 1 is disclosed.

A non-transient storage medium according to claim 9 is also disclosed.

A color similarity evaluation device according to claim 10 is also disclosed.

Brief description of the drawings

The following figures provide preferred embodiments to illustrate the description and are not to be construed as limiting the scope of the invention.

Figure 1: Schematic representation of the relationships between the color difference results using AE ⁰⁰ and the corresponding value obtained per modulus. In dark gray the values of AE ⁰⁰ above the threshold and in light gray the values below or equal to the threshold. The horizontal line represents a suggested threshold for the module.

Figure 2: Schematic representation of the relationships between the results of the color differences using AE ⁰⁰ and the corresponding value obtained by the angle. In dark gray, the AE ⁰⁰ values above the threshold and in light gray the values below or equal to the threshold. The horizontal line represents a suggested threshold for the angle.

Figure 3: Schematic representation of the relationships between the results of color differences using AE ⁰⁰ and the corresponding value obtained for lightness. In dark gray, the AE ⁰⁰ values above the threshold and in light gray the values below or equal to the threshold. The horizontal line represents a suggested threshold for lightness.

Figure 4: Schematic representation of the module, angle and brightness thresholds for each AE_00 threshold.

Figure 5: Schematic representation of a comparison between the optimal and predicted threshold ratio of modulus, angle and luminosity with respect to the AE00 threshold.

Figure 6: Schematic representation of the histogram of color differences that can be filtered by each module.

Figure 7: Schematic representation of the block diagram of the cascaded embodiment.

Figure ⁸ : Schematic representation of the block diagram of the parallel implementation.

Figure 9: Schematic representation of the processing time for the different strategies, taking into account two decision thresholds (vertical lines). Figure 9(a) refers to serial processing of pixels in an image and Figure 9(b) refers to parallel processing. For example, when applying Td ¹ the decision is made by the module M ¹ , while for Td ² only the module M ³ can make a decision.

Figure 10: Schematic representation of image processing possibilities taking advantage of multi-core CPUs. The Fast Color Decision (FCD) module represents one of the two proposed strategies.

Detailed description

The thresholds indicated in Figures 1 through 3 are only valid for AE ⁰⁰ values of 3, 5, and 10. For other metric values, new thresholds can be identified. Models for modulus, angle, and luminosity were derived to avoid empirical adjustment of these thresholds. These were obtained by training regression models from the data derived from the 5 million calculated color differences. For the data set, the results and thresholds were computed for AEooe {0.30} with a step of 0.05. These values were selected because colors with a color difference of 0 are visibly imperceptible and colors with a color difference of 30 are extremely different. The threshold relationships obtained are presented in Figure 4. It can be seen that the relationships for modulus and angle are non-linear, but are approximately linear for luminosity. To obtain the property models and validate the approach, it is decided to train a random forest [ ² ] for module and angle and a linear regression for luminosity. Models were trained using a cross-validation approach, and models with the best fit had mean squared errors of 0.00675, 0.0333, and 0.00602 for angle, modulus, and luminosity, respectively. Ideally, these errors should be as close to 0 as possible, but for the scope of this disclosure, the errors obtained were considered acceptable since the data used for training was already regularized.

This regularization was introduced into the training data to shift the predicted values so that possible errors would not influence the final decision. Although this results in predicted thresholds that are different from the optimal values, they are guaranteed to always be above the corresponding optimal values and therefore always applicable. This results in performance deterioration, while ensuring corrected predictions. The comparison of the predicted thresholds using regularized and non-regularized data is shown in Figure 5, where the black lines correspond to the regularized predicted thresholds and the dotted lines correspond to the non-regularized data. Analyzing the charts, it becomes clear that, in the universe of stocks studied, the predictions are always out of date.

With the derived models it is possible to obtain, for any value of AE ⁰⁰ , modulus, angle and lightness thresholds, so that any two colors whose difference exceeds these thresholds can be classified as having an AE ⁰⁰ color difference greater than the specified value. This is very advantageous, as it can be mapped into a structure that effectively trims the computations, resulting in a significant performance gain without compromising the properties of AE ⁰⁰ . Each threshold can be used to reduce the number of AE ⁰⁰ calculations required for a given application by setting a limit for AE ⁰⁰ . However, the color combinations that each threshold cuts are not the same. Figure ⁶ represents a histogram of the number of color differences that each module can filter for the three values of AE ⁰⁰ considered above. The histograms show that a very small number of color differences can be filtered exclusively by module, angle or lightness. Although this number increases with the AE00 threshold, it is still very low. It is also evident that a very large number of color differences can be filtered by combining module, angle or brightness. This number varies inversely to the AE ⁰⁰ threshold and the decrease is compensated by increasing differences filtered by only two of the properties, especially modulus-luminosity and modulus-angle, but also those that can be decided only by computing the full AE ⁰⁰ metric . Therefore, different strategies can be defined to effectively validate whether or not it is necessary to compute the full color difference formula. Next, a set of possible realizations is proposed and analyzed.

Based on the described concepts and relationships, different strategies can be conceived to make a faster decision on the similarity of colors based on the AE ⁰⁰ metric. In this section, two embodiments are presented and discussed: ( ¹ ) cascaded (ie computed successively) filter models; (2) filtering models deployed in parallel. In addition, an embodiment that can be applied to large varieties of color differences to be computed is discussed, specifically in image processing, where initial proposals for color difference calculations can be used, specifically taking advantage of multi-core hardware. widely spread.

The first proposed strategy for efficient color decision based on AE ⁰⁰ is to cascade (i.e. compute in succession) the filter modules as illustrated in Figure 7. In each module in the cascade, simpler calculations are made to calculate the module, angle or luminosity and the evaluation is carried out. The full AE ⁰⁰ metric is only computed as a last resort; therefore, it will always be the last module in the cascade.

Each module in the cascade is meant to make a decision: if the colors are different, then cascade output; or, uncertainty exists and processing is passed to the next module in the cascade. Therefore, it is expected that it is desirable that the fastest modules, with the least computational weight, are executed first, preferably by pruning as many pixels as possible. However, determining the most appropriate cascading structure, that is, identifying the order of operations, is not necessarily defined by the speed of execution. It is important to note that the cascading structure may depend on the hardware architecture. Furthermore, the threshold value of AE ⁰⁰ to be considered may influence the best cascade structure. Therefore, optimized embodiments are described below.

The cascading approach aims to sequentially perform operations that reduce the number of color pairs that require the calculation of AE ⁰⁰ . Since the initial blocks of the cascade cannot decide on the dissimilarity of colors and AE ⁰⁰ must ultimately be computed, the total number of computations would in fact be larger than just for the computation of AE ⁰⁰ . Therefore, the advantages of the cascading strategy derive from a greater number of colors filtered by previous blocks.

An alternative strategy to cascading is proposed, in which the different blocks are executed in parallel, as illustrated by the block diagram in Figure 8. The underlying assumption is that, in this case, the maximum time to analyze a difference in color does not exceed AE ⁰⁰ time. Although this may not be exactly true due to synchronization and management overhead, this difference can be considered almost negligible. These embodiments for the cascaded and parallel cases are illustrated in Figure 9(a) and Figure 9(b) respectively. Obviously, the advantage of the parallel strategy is directly related to the availability of means of parallel execution, for example, the number of CPU cores available.

Although this approach is robust, it is not as straightforward to apply as the waterfall approach. In fact, several implementation details must be taken into account or jeopardize possible profits.

The described strategies are generic for a decision on color similarity. As stated above, the advantages of the parallel approach are most apparent when multi-core CPUs and distributed processing can be used.

When a large variety of colors is to be processed, for example an image, ie compared to one or more reference colors, this data may be processed sequentially. In this case, each color would be individually and sequentially passed as input along with the reference color and the desired threshold. In other words, a couple of colors at a time. This is illustrated in Figure 10(a). The calculation of the color difference itself could be done by any of the strategies described in the previous subsections. When multiple cores are available, the data to be processed can be partitioned and each of these blocks can be passed as input for processing, as illustrated in Fig. 10(b).

The order of the blocks in the cascading strategy (that is, the order in which they are calculated in succession) influences the result of the overall process, more specifically the time reduction that can be obtained compared to the calculation of the AE metric ⁰⁰ . Block AE ⁰⁰ will be the last in the cascade, but it is still necessary to identify what the remaining blocks should be and their order. Given the three exposed properties, there are 15 possible cascading structures, as identified in Table 1. Each of these possibilities performs a different pruning on color differences. Furthermore, the complexity of the cascade will depend on the constraints imposed, namely the required decision threshold AE ⁰⁰ .

Table 1: Identification of possible cascading structures. Even if it is not represented, block AE ⁰⁰ will always be the last module in the cascade.

To evaluate the possibilities of the cascade, a simple segmentation algorithm is implemented that, given an image, a color and a threshold of AE00, produces a new image that highlights the regions where the colors are below the specified threshold. Each structure was evaluated using a generic color bank as specified in [5] and images from the Caltech 256 dataset [1]. A subset of 4K resolution images from the RAISE data set [ ⁶ ] and a small set of computer generated ^8K images were also used. Each instance of the color bank was compared with the color of each pixel of the images of the three subsets. Furthermore, the evaluation was performed on three different computer architectures to understand the impact on the cascading structure. Specifically, the following architectures were used (1) a desktop computer equipped with an Intel H 4471 processor at 3.5Ghz; (2) a Raspberry Pi 2, model B; (3) a Samsung Galaxy S ⁶ smartphone. The algorithms were implemented in C++ for the first two environments and in Java for the smartphone. No explicit GPU acceleration was used in these experiments, and CPU time and cycles allocated to the process were measured.

The results presented in Table 2 clearly show the advantage of the waterfall approach as an alternative to full computation of the AE ⁰⁰ formula for a desktop computer. Regardless of the filter blocks, modulus, angle, or luminosity, their order, and the underlying AE ⁰⁰ threshold, the cascading strategy always achieved significant performance gains. This is true when measuring processing time as well as CPU cycles. Since there is a clear relationship between the number of CPU cycles and the total processing time, in the following results CPU cycles will only be presented when justified to convey additional information.

Increasing the threshold of AE ⁰⁰ implies some changes in the classification and the overall gain decreases. However, the time reductions that the cascading strategy allows are still very significant (above 70%) and the results show a consistency of the cascading structure. Although there are some changes in the classification, namely LMA-LAM and M-ML, the differences in relative gain are almost negligible. The results show that, in general, higher gains are obtained when the module is the first block in the cascade. This is consistent with the fact that (1) it is a light operation; (2) does not require a color conversion; (3) allows filtering a large number of color differences, as can be seen in figure ¹ . This is especially true for an AE ⁰⁰ threshold of 3 (table ² ), where the best results were obtained for a cascade with only the modulo filter block. Although the module, as a filter block, provides large computational gains by itself, in general better results were obtained by adding other blocks; this is due to the simplicity of the formulas compared to AE ⁰⁰ and is indicative of some filtering complementarity between module, angle and luminosity, as shown by the histograms in figure ⁶ . As expected, using luminosity as the first block results in lower time gains, due to its more complex formulation (compared to modulus and angle) and the need for color space conversion. It is also interesting to note that the highest gain is obtained with the module-angle combination; adding lightness as the third block results in a slightly lower gain.

The results of the experiments on a Raspberry Pi 2, model B are consistent with the results obtained on a desktop computer, with relative gains of the same order, despite the more limited resources. Also, as in the desktop computer, some modifications are observed in the classification of the cascading structure that are due to almost negligible differences in the relative gain. It is remarkable the fact that, as in the case of the desktop computer, the module presents a high impact and for AE ⁰⁰ = 3 it provides the best result by itself.

The classifications for Android have some differences with respect to the previous cases, especially with the desktop computer, since the classification that begins with MA is followed by AM, M or AML, or that begins with AML and is followed by MAL, AM, etc However, the gains are very significant and only slightly lower (between 65% and 89%) than in previous cases. Differences in computer architectures undoubtedly contribute to this, but also the requirements imposed on implementation languages. For the desktop computer and the Raspberry the blocks were implemented in C++ and for Android Java was used.

Table 2: Desktop cascade performance for AE ⁰⁰ = 3 for the Caltech 256 data set. Results are ordered inversely of their computational cost.

The proposed embodiments allowed very significant gains to be made on all used hardware platforms. Slightly lower gains were achieved using Android, which is not agnostic not only to platform differences, but also to the implementation language itself. However, the differences in relative gain are minimal. Although the gains depend on the desired decision bound on the result of AE ⁰⁰ , they are very high for useful values of the metric; Values greater than AE ⁰⁰ will have less and less entropy, taking into consideration that values greater than 7 represent two very different colors.

The best structure for the cascading strategy varied slightly, but in general a small set of possibilities can be identified, to be used horizontally with very small decreases in relative gain. Specifically, for execution on a desktop computer, it is possible to identify the structure comprised by the modulo-angle filter blocks followed by the AE ⁰⁰ calculation block as the one that provides the best results or only slightly below. These same conclusions can be applied to the cases of the Raspberry Pi and the smartphone. As before, the structure comprised of the modulo-angle filter blocks followed by the AE ⁰⁰ block can be identified as a good compromise for a structure that can be applied regardless of the range of target values of AE ⁰⁰ . For these two cases, the identified cascade structure not only provides the best result more frequently, but the differences are nonetheless very small.

Although AE ⁰⁰ values greater than ^{1 0} are not taken into account, it should be noted that colors with an AE ⁰⁰ distance greater than this value are considered noticeably different. However, the insights resulting from the described analysis and the proposed models (see Figure 4) allow the implementation of the cascade in such a way that it can be adapted to different decision thresholds and hardware platforms known a priori and given a context passed as input. .

The filtering properties analyzed, modulus, angle and lightness, allowed very high processing gains using different strategies and hardware platforms. However, other properties can be analyzed and their advantages studied. Furthermore, it is possible to use a hybrid strategy that combines the cascading and parallel strategy.

The term "comprising", whenever used in this document, is intended to indicate the presence of mentioned features, integers, steps or components, but not to exclude the presence or addition of one or more features, integers, steps, components or groups thereof. In the figures, flowcharts of particular embodiments of the methods disclosed herein are depicted. The flowcharts do not represent any particular means, instead the flowcharts illustrate the functional information that a person skilled in the art needs to carry out said methods required according to the present disclosure. Those skilled in the art will appreciate that, unless otherwise indicated herein, the particular sequence of steps described is merely illustrative and can be varied without departing from the disclosure. Therefore, unless otherwise indicated, the steps described are so unordered that, where possible, the steps can be performed in any convenient or desirable order.

It should be appreciated that certain embodiments of the disclosure described herein may be embodied as code (eg, an algorithm or software program) that resides in firmware and/or on usable media. by a computer having control logic to enable execution on a computer system having a computer processor, such as any of the servers described herein. Such a computer system typically includes storage memory configured to provide the output of code execution that configures a processor based on execution. The code may be arranged as firmware or software, and may be organized as a set of modules, including the various modules and algorithms described herein, such as discrete code modules, function calls, procedure calls, or objects in a framework environment. object oriented programming. If implemented using modules, the code may comprise a single module or a plurality of modules that operate in cooperation with one another to configure the machine on which it runs to perform associated functions, as described herein.

Claims (10)

1. A computer-implemented method for evaluating the similarity of colors in image processing between a first color of an image to be processed and a second color, which is a reference color, said method comprising the following steps: apply one or more color difference filters to said first and second colors, selected from: a modulus filter obtained by calculating the modulus of the difference vector between said first and second colors represented as 3D coordinates in an RGB color space, an angle filter obtained by calculating the angle between the two vectors of the first and second colors represented as 3D coordinates in an RGB color space, a lightness filter obtained by calculating the difference in lightness between the first and second colors in the CIELAB color space; determining that the first and second colors are dissimilar if at least one of said one or more applied filters is above a predetermined threshold for each of the calculated modulus, angle, and luminosity filters; otherwise, calculating a CIEDE2000 AEoo color difference and determining that the first and second colors are dissimilar if the calculated CIEDE2000 AEoo color difference is above a predetermined CIEDE2000 AEoo color difference threshold; otherwise, determining that the first and second colors are similar; wherein regression models that are pretrained on a plurality of precomputed CIEDE2ooo color differences obtained from an image data set are applied to determine a relationship between each of the filter threshold values and the values color difference threshold value CIEDE2ooo AEoo, and obtaining from said relations the predetermined thresholds for each of the modulus filter, angle filter and luminosity filter for the predetermined color difference threshold CIEDE2ooo which is set automatically. that two colors are considered perceptually different. A computer-implemented method according to the preceding claim, wherein the step of applying one or more color difference filters comprises applying the modulus filter and the angle filter. The computer-implemented method of claim 2, wherein the step of applying one or more color difference filters comprises applying the modulus filter and the angle filter, applied in succession, thereby applying the angle filter after applying the modulo filter. The computer-implemented method of claim 3, wherein the step of applying one or more color difference filters comprises applying the modulus filter, the angle filter, and the luminosity filter, applied in succession, by applying, thus, the luminosity filter after the angle filter, and applying the angle filter after applying the module filter. The computer-implemented method of claim 2, wherein the step of applying one or more color difference filters comprises applying the modulus filter, the angle filter, and the luminosity filter, in parallel. A computer-implemented method according to the preceding claim, wherein the step of applying one or more color difference filters comprises applying the modulus filter, the angle filter, and the luminosity filter, in parallel such that the application of said one or more filters is interrupted when at least one of the applied filters is above a predetermined threshold for said filter. A computer-implemented method according to any of the preceding claims, wherein the predetermined threshold for the luminosity filter is pre-obtained from the AEoo threshold using a pretrained linear regression model between the AEoo threshold and the predetermined threshold for the lightness filter. A computer-implemented method according to any one of the preceding claims, wherein the predetermined threshold for the modulus filter and/or the predetermined threshold for the angle filter is pre-derived from the AEoo threshold using an algorithm model of pretrained random forest algorithm machine learning between the AEoo threshold and the default threshold for the modulus filter and/or the default threshold for the angle filter. Non-transient storage media including program instructions which, when executed by a computer, cause the computer to carry out the method according to any one of claims 1 to 8. A color similarity evaluation device comprising the non-transient storage means according to claim 9 and a data processor arranged to execute said instructions.