A new Image Similarity Metric for a Perceptual and Transparent Geometric and Chromatic Assessment

arXiv:2601.19680 · cs.CV · Submitted 2026-01-27 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "A new Image Similarity Metric for a Perceptual and Transparent Geometric and Chromatic Assessment".

Jane: Here's the gist of the paper. The gist A new perceptual metric composed of two terms, one for texture dissimilarity using Earth Mover’s Distance and another for chromatic dissimilarity in Oklab space, is proposed to outperform state-of-the-art metrics by providing a transparent and justified similarity assessment How it works The proposed metric, EDOKS, is composed of two terms that evaluate different aspects of image dissimilarity The first term evaluates the dissimilarity between the textures of two images using Earth Mover’s Distance (EMD) This texture term is calculated by dividing the image into non-overlapping patches and applying a Gabor filter dictionary to each patch to obtain texture signatures S The EMD is then calculated between these two texture signatures, where the goal is to measure the minimum amount of work to fill the holes in one distribution using the ground of another The second term evaluates chromatic dissimilarity between two images in the Oklab perceptual color space The color dissimilarity is calculated by converting both RGB images to the Oklab color space and then calculating the perceptual distance ∆E between all pixels of the two images This distance is defined as a straightforward Euclidean distance between the components of two color points in Oklab space, resulting in a dissimilarity value between 0 and 1 The final EDOKS index is proposed as a combination of these two terms It is defined by the equation EDOK(X, Y) = αEMD(S X, S Y) + (1 − α)OK(Xˆ, Yˆ) where OK is the dissimilarity term derived from the Oklab distance and S X and S Y are the texture signatures The weight parameter α is a value between 0 and 1 that allows for weighting the two terms based on context The final similarity term, EDOKS (X, Y) = 1/EDOK(X, Y) + c (5), is presented as the reciprocal of the dissimilarity term to yield a similarity score Key Components and Methodology The texture dissimilarity term relies on the work of Rubner et al.

16: , which proposes an image retrieval method based on EMD, solving the Monge-Kantorovich transport problem Texture signatures are obtained by applying a dictionary of Gabor filters with four scales and six orientations to each patch, resulting in a collection of twenty-four total filters for each patch The signature S is then obtained by clustering the energies using the Meng-Hee Heng algorithm, which automatically finds the suitable number of clusters k The color dissimilarity term utilizes the Oklab color space because it is designed to ensure uniformity and perceptual linearity in hue The perceptual distance ∆E(p1, p2) between two color points p1 and p2 in Oklab space is calculated via the straightforward Euclidean distance between the components of p1 and p2 This allows for a faithful comparison of color differences according to human perception Performance and Validation Experiments were conducted on the Berkeley-Adobe Perceptual Patch Similarity (BAPPS) Dataset The results demonstrated that EDOKS behaves consistently with the perceptual similarity of color and textures between images, a behavior not always guaranteed by state-of-the-art similarity metrics When compared against other low-level FR methods, EDOKS was found to be the closest similarity metric to human perceptual similarity, while it is also close to the results of other deep metrics The statistical comparison on the JND dataset showed that EDOKS significantly outperforms other SOTA low-level metrics The Spearman Rank Order Correlation Coefficient (SROCC) for EDOKS was 0.537, which is higher than the scores obtained by other low-level metrics like PSNR (0.390) and SSIM (0.365) Furthermore, the ablation study confirmed that the combination of both terms yields a better result than using either term individually Interpretability A key advantage of EDOKS is its transparency, as it provides visual explanations to support the calculated score The metric provides maps of interest that highlight areas with significant differences between two images, ensuring transparency in its use Figure 7 illustrates examples of EDOKS’s ability to identify discriminatory areas with the greatest difference between two images X and Y, providing maps of interest based on differences in shapes and colors The EMD difference column highlights the points where the two images differ in shape, while the OK heatmap column highlights the color difference ∆E(Xˆi, j, Yˆi, j) between all pixels of the two images in the Oklab space This visual evidence helps users evaluate their images beyond simply using numerical scores The analysis shows that EDOKS is not influenced by distractors because it never highlights the subject of the image and does not present any type of distortion In conclusion, EDOKS is recommended for use because it is a more valid perceptual metric compared to others, especially in the presence of geometric distortion Unlike other similarity indices based on black-box models, such as LPIPS, EDOKS is interpretable by analyzing the individual terms EMD and OK The authors suggest that future developments may involve modifying parameters such as patch size p, Gabor filtering, clustering algorithm or distance used in OK The implementation of the proposed similarity metric is publicly available at https://github.com/antdimarino/EDOKS Page 3 --- Antonio Di Marino and Vincenzo Bevilacqua are with the Institute for High-Performance Computing and Networking (ICAR) - National Research Council (CNR), Naples, Italy, and also with the University of Naples Federico II, Department of Electrical Engineering and Information Technology, via Claudio 21, Naples, 80125, Italy (email: antonio.dimarino@icar.cnr.it; vincenzo.bevilacqua@icar.cnr.it) The official implementation of our metric is publicly available at https://github.com/antdimarino/EDOKS Page 4 --- Fig.

Paper summary: Tom: So we're looking at this paper called "A new Image Similarity Metric for a Perceptual and Transparent Geometric and Chromatic Assessment." Basically, they say that current state-of-the-art image similarity metrics aren't really perceptual anymore, especially when you have texture distortion mixed in.

Jane: That makes sense. So what is this new metric that they propose? What’s the main idea behind their approach?

Tom: They propose a new metric built from two terms. The first term handles the texture dissimilarity using Earth Mover's Distance, and the second term looks at chromatic dissimilarity in a color space called Oklab.

Lu: I read that they use Earth Mover's Distance to evaluate texture dissimilarity between two images, which is based on solving the Monge-Kantorovich transport problem <ref:2601.19680#pg2>.

Meng: Okay, so for the first part, they divide the image into patches and use a Gabor filter dictionary to get texture signatures S, and then they calculate the EMD between those signatures. That sounds computationally heavy.

Tom: It is complex, but that’s how they define the texture term. Then there's this second term which looks at color similarity in Oklab space by calculating a perceptual distance Delta E between all pixels in that space.

Jane: So they are combining these two different types of comparison—texture and color—to get one overall similarity score? What is the big claim here?

Tom: They tested this on a non-traditional dataset called Berkeley-Adobe Perceptual Patch Similarity, which has a lot of complex distortions in shapes and colors. And they showed their metric outperforms the state of the art, especially when those shape distortions are present.

Lu: That's interesting because it confirms its greater perceptiveness even with geometric issues <ref:2601.19680#pg2>. They also mention that deep black-box metrics can be accurate but they just give you a score without explaining the differences.

Meng: So, the value of this metric isn't just the number, it’s the explanation behind it? That transparency is something I need to see if I want to use it in production.

Jane: Exactly. They say their metric provides visual explanations that highlight areas with significant differences between two images, showing you where the shapes and colors really diverge.

Paper summary: Tom: So they are giving us maps of interest, which show exactly where the two images differ based on shape and color in Oklab space. That makes it much more transparent than some other methods we've seen <ref:2601.19680#pg3>.

Lu: I read that the texture term uses a dictionary of Gabor filters with four scales and six orientations for each patch, resulting in twenty-four total filters for each patch. Then they use the Meng-Hee Heng algorithm to cluster the energies and find the number of clusters k automatically <ref:2601.19680#pg3>.

Meng: Twenty-four filters per patch sounds like a lot of work for real traffic, but if it’s stable, maybe it works. What about the color space part? How does Oklab help?

Jane: The Oklab space is chosen because it's designed to ensure uniformity and perceptual linearity in hue. They calculate the perceptual distance Delta E as a straightforward Euclidean distance between the components of two color points in that space <ref:2601.19680#pg3>.

Tom: And they combine these two parts using an equation where alpha is a weight between zero and one, allowing you to weigh the texture term against the Oklab distance term <ref:2601.19680#pg3>. The final similarity score is presented as the reciprocal of this dissimilarity term.

Meng: So, if we’re thinking about deployment at scale, what does that mean for computational cost? Is it something you can run fast enough?

Tom: Well, they state their metric outperforms other low-level FR methods significantly on the Berkeley-Adobe Perceptual Patch Similarity dataset <ref:2601.19680#pg2>. The Spearman Rank Order Correlation Coefficient for their EDOKS index is zero point five three seven, which is higher than PSNR at zero point three nine zero and SSIM at zero point three six five on the JND dataset.

Jane: So, for someone just looking at the numbers, it shows a better correlation with what humans actually perceive in those complex distortions than standard measures like SSIM or PSNR do.

Lu: And they also did an ablation study showing that combining both terms gives a better result than using either term by itself <ref:2601.19680#pg3>. It seems the combination is key to getting that perceptual lift.

Meng: I gotta ask about the limitations mentioned in the paper. What are they admitting this metric doesn't do?

Tom: They point out that although deep black-box metrics can be very accurate, they don't provide any visual explanations of why two images are different <ref:2601.19680#pg2>. That’s the main reason they built it—to add that transparency.

Paper summary: Jane: So, if you only look at the numerical score from this paper, you miss out on the context of *why* it scored that way, which is what makes this metric different from some of those black-box models.

Tom: It’s about moving beyond just a similarity number to actually understanding the differences in shape and color visually <ref:2601.19680#pg3>. So, this approach is designed for situations where geometric distortion is a big deal.

Lu: The authors suggest future work could involve modifying parameters like the patch size p, the Gabor filtering, or even changing the distance used in the Oklab calculation <ref:2601.19680#pg3>. They are open to tweaking the components.

Jane: So what does this mean for how we use image similarity tools going forward? It seems they are pushing for metrics that aren't just mathematically sound but also truly perceptual and transparent.

Tom: They’re proposing a way to get a score that is justified by showing you the visual evidence, which makes it a more valid tool than just relying on one single number.

Meng: I guess for those of us in production, the cost of generating these detailed maps is something we have to weigh against the accuracy gain they’re showing here <ref:2601.19680#pg3>. It's a trade-off between cost and that perceptual improvement.

Jane: So, to wrap up this look at "A new Image Similarity Metric for a Perceptual and Transparent Geometric and Chromatic Assessment," they propose EDOKS as a way to get better, more transparent similarity scores by combining texture dissimilarity with Oklab color space analysis.

Tom: It’s about making the assessment of images more reliable when shapes are distorted because it gives you the visual proof behind the score.

Lu: The idea of separating texture and chromatic components and then blending them using a tunable weight alpha is a solid way to build a composite metric <ref:2601.19680#pg3>.

Meng: It seems like this is moving us away from just trusting the output of massive black-box models without seeing the underlying reasons for their decisions.

Jane: So, this paper suggests that for complex image comparison, we need metrics that are not only accurate but also explain their reasoning through visual maps.

Conclusion: Tom: So we're wrapping up this look at "A new Image Similarity Metric for a Perceptual and Transparent Geometric and Chromatic Assessment." Basically, they've put together this new way to compare images that tries to be both perceptual and transparent.

Jane: That sounds like a big title. So what exactly is this metric doing? I want the simple version, not the math stuff right now.

Lu: It’s combining two things—texture similarity and color similarity—and they use Oklab space for the colors, which is supposed to be really good for how humans see hue.

Meng: So it’s trying to capture both what you feel when you look at the surface and what you feel about the colors at the same time. I need to know if this actually works in a messy production environment.

Lalam: The core idea is that current state-of-the-art metrics often fail when shapes are distorted, and this one specifically tries to fix that by adding a texture component using Earth Mover's Distance.

Tom: Yeah, I saw the results on the Berkeley-Adobe Perceptual Patch Similarity dataset. They showed it actually beats the existing state of the art when there are shape distortions around. That’s a concrete win on performance against these complex inputs.

Jane: It sounds like they’re saying that if you only look at color or just texture, you miss what's going on when both are messed up together. What about that transparency thing they mentioned?

Lu: They aren't just giving you a score number; they provide visual explanations. They show maps of interest that highlight exactly where the shape and color differences are most significant between two images.

Meng: Visual explanations are good, but what does that cost? If I have millions of images to process, generating these detailed maps might be too slow for what I need in production.

Tom: That’s the trade-off they’re making. They admit deep black-box metrics can be accurate but they don't explain *why* the score is what it is. This metric gives you that explanation, which makes it more justifiable than just a number from a black box model.

Jane: So for someone who just uses these tools, this means you get to see the visual evidence supporting the similarity score instead of just trusting a random number. That’s a big difference in how you use it daily.

Lu: They’re open to tweaking things like the patch size or the Gabor filters later on, so they're building something flexible that can adapt to different kinds of distortion.

Lalam: The implication is that for applications dealing with geometric distortions—like comparing objects under bad lighting or in a noisy environment—this composite approach offers a more valid way to assess similarity.

Tom: Yeah, it seems like the authors are pushing for metrics that aren't just mathematically sound but actually reflect human perception when things get distorted.

Antonio Di Marino, Vincenzo Bevilacqua, Emanuel Di Nardo, Angelo Ciaramella, Ivanoe De Falco, Giovanna Sannino

Institute for High-Performance Computing and Networking (ICAR) - National Research Council (CNR) · University of Naples Federico II

cs.CV

Submitted: 2026-01-27

Updated: 2026-01-27

Journal ref: Discover Artificial Intelligence 6, 1536 (2026)

DOI: 10.1007/s44163-026-02424-4

Code: https://github.com/antdimarino/EDOKS

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 72/100

The gist: The gist A new perceptual metric composed of two terms, one for texture dissimilarity using Earth Mover’s Distance and another for chromatic dissimilarity in Oklab space, is proposed to outperform

Key concepts

Earth Mover's Distance (EMD): EMD measures the minimum amount of work required to transform one image's texture distribution into another. It works by dividing images into patches and using Gabor filters to create texture signatures. The EMD then calculates the effort needed to move these texture distributions from one image to another, quantifying how different their textures are.
Oklab Perceptual Color Space: This is a specialized color space designed for human perception, aiming for uniformity and linearity in hue. By converting images from RGB to Oklab, the metric calculates the perceptual distance between colors using Euclidean distance on this space. This ensures that the calculated color difference more closely matches how humans perceive changes in color.
EDOKS Index: The final EDOKS index combines texture dissimilarity (EMD) and chromatic dissimilarity (Oklab). It is weighted by a parameter $\alpha$ to balance these two aspects. The resulting similarity score is the reciprocal of the total dissimilarity, yielding a value where higher scores indicate greater perceptual similarity between two images.

Summary

The gist A new perceptual metric composed of two terms, one for texture dissimilarity using Earth Mover’s Distance and another for chromatic dissimilarity in Oklab space, is proposed to outperform state-of-the-art metrics by providing a transparent and justified similarity assessment

How it works

The proposed metric, EDOKS, is composed of two terms that evaluate different aspects of image dissimilarity The first term evaluates the dissimilarity between the textures of two images using Earth Mover’s Distance (EMD) This texture term is calculated by dividing the image into non-overlapping patches and applying a Gabor filter dictionary to each patch to obtain texture signatures S The EMD is then calculated between these two texture signatures, where the goal is to measure the minimum amount of work to fill the holes in one distribution using the ground of another

The second term evaluates chromatic dissimilarity between two images in the Oklab perceptual color space The color dissimilarity is calculated by converting both RGB images to the Oklab color space and then calculating the perceptual distance ∆E between all pixels of the two images This distance is defined as a straightforward Euclidean distance between the components of two color points in Oklab space, resulting in a dissimilarity value between 0 and 1

The final EDOKS index is proposed as a combination of these two terms It is defined by the equation EDOK(X, Y) = αEMD(S X, S Y) + (1 − α)OK(Xˆ, Yˆ) where OK is the dissimilarity term derived from the Oklab distance and S X and S Y are the texture signatures The weight parameter α is a value between 0 and 1 that allows for weighting the two terms based on context The final similarity term, EDOKS (X, Y) = 1/EDOK(X, Y) + c (5), is presented as the reciprocal of the dissimilarity term to yield a similarity score

Key Components and Methodology

The texture dissimilarity term relies on the work of Rubner et al. [16], which proposes an image retrieval method based on EMD, solving the Monge-Kantorovich transport problem Texture signatures are obtained by applying a dictionary of Gabor filters with four scales and six orientations to each patch, resulting in a collection of twenty-four total filters for each patch The signature S is then obtained by clustering the energies using the Meng-Hee Heng algorithm, which automatically finds the suitable number of clusters k

The color dissimilarity term utilizes the Oklab color space because it is designed to ensure uniformity and perceptual linearity in hue The perceptual distance ∆E(p1, p2) between two color points p1 and p2 in Oklab space is calculated via the straightforward Euclidean distance between the components of p1 and p2 This allows for a faithful comparison of color differences according to human perception

Performance and Validation

Experiments were conducted on the Berkeley-Adobe Perceptual Patch Similarity (BAPPS) Dataset The results demonstrated that EDOKS behaves consistently with the perceptual similarity of color and textures between images, a behavior not always guaranteed by state-of-the-art similarity metrics When compared against other low-level FR methods, EDOKS was found to be the closest similarity metric to human perceptual similarity, while it is also close to the results of other deep metrics

The statistical comparison on the JND dataset showed that EDOKS significantly outperforms other SOTA low-level metrics The Spearman Rank Order Correlation Coefficient (SROCC) for EDOKS was 0.537, which is higher than the scores obtained by other low-level metrics like PSNR (0.390) and SSIM (0.365) Furthermore, the ablation study confirmed that the combination of both terms yields a better result than using either term individually

Interpretability

A key advantage of EDOKS is its transparency, as it provides visual explanations to support the calculated score The metric provides maps of interest that highlight areas with significant differences between two images, ensuring transparency in its use Figure 7 illustrates examples of EDOKS’s ability to identify discriminatory areas with the greatest difference between two images X and Y, providing maps of interest based on differences in shapes and colors The EMD difference column highlights the points where the two images differ in shape, while the OK heatmap column highlights the color difference ∆E(Xˆi, j, Yˆi, j) between all pixels of the two images in the Oklab space This visual evidence helps users evaluate their images beyond simply using numerical scores The analysis shows that EDOKS is not influenced by distractors because it never highlights the subject of the image and does not present any type of distortion

Conclusion

In conclusion, EDOKS is recommended for use because it is a more valid perceptual metric compared to others, especially in the presence of geometric distortion Unlike other similarity indices based on black-box models, such as LPIPS, EDOKS is interpretable by analyzing the individual terms EMD and OK The authors suggest that future developments may involve modifying parameters such as patch size p, Gabor filtering, clustering algorithm or distance used in OK The implementation of the proposed similarity metric is publicly available at https://github.com/antdimarino/EDOKS

--- Page 3 ---

Antonio Di Marino and Vincenzo Bevilacqua are with the Institute for High-Performance Computing and Networking (ICAR) - National Research Council (CNR), Naples, Italy, and also with the University of Naples Federico II, Department of Electrical Engineering and Information Technology, via Claudio 21, Naples, 80125, Italy (email: antonio.dimarino@icar.cnr.it; vincenzo.bevilacqua@icar.cnr.it) The official implementation of our metric is publicly available at https://github.com/antdimarino/EDOKS

--- Page 4 ---

Fig.

Improvements for AI systems

No black-box similarity metrics like LPIPS will be replaced by EDOKS, which provides visual explanations to support the calculated score, making the similarity assessment transparent and justified. This allows developers to understand precisely what aspects (texture or color) of a distortion caused the similarity score.
The AI system can accurately evaluate complex distortions in shapes and colors because EDOKS is based on the integration of low-level handcrafted features, specifically using Earth Mover’s Distance for texture dissimilarity and Oklab space for chromatic dissimilarity.
The system will be more robust against various types of visual artifacts, such as shape distortions and color distortions, because it utilizes both a texture term (EMD) and a color term (Oklab similarity), unlike metrics that only focus on luminance or structural information.
The AI can distinguish between image pairs with high confidence when the distortion involves both shape and color, as shown in Figure 8 where EDOKS highlights differences in both shape (EMD difference) and color (OK heatmap).