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

TL;DR

This work addresses the mismatch between traditional image quality metrics and human perception by proposing EDOKS, a white-box perceptual similarity metric that jointly quantifies texture dissimilarity via Earth Mover's Distance on Gabor-based texture signatures and color dissimilarity via Euclidean distance in the perceptually uniform Oklab space. The two terms are combined as $ ext{EDOK}(X,Y)= ext{ED}^{EMD}+ ext{ED}^{OK}$ with a tunable weight $ ext{alpha}$, and the final similarity is $ ext{EDOKS}(X,Y)=rac{1}{ ext{EDOK}(X,Y)+c}$. Evaluations on the BAPPS and LIUK4-v2 datasets reveal that EDOKS aligns more closely with human judgments than standard low-level metrics and remains competitive with deep metrics while offering interpretable heatmaps that highlight the regions responsible for the score. Ablation studies confirm the necessity of both terms, and qualitative interpretability analyses demonstrate how texture and color differences jointly shape the perceptual assessment. The work contributes a practical, transparent IQA tool with broad applicability to image generation and distortion evaluation, along with publicly available code.

Abstract

In the literature, several studies have shown that state-of-the-art image similarity metrics are not perceptual metrics; moreover, they have difficulty evaluating images, especially when texture distortion is also present. In this work, we propose a new perceptual metric composed of two terms. The first term evaluates the dissimilarity between the textures of two images using Earth Mover's Distance. The second term evaluates the chromatic dissimilarity between two images in the Oklab perceptual color space. We evaluated the performance of our metric on a non-traditional dataset, called Berkeley-Adobe Perceptual Patch Similarity, which contains a wide range of complex distortions in shapes and colors. We have shown that our metric outperforms the state of the art, especially when images contain shape distortions, confirming also its greater perceptiveness. Furthermore, although deep black-box metrics could be very accurate, they only provide similarity scores between two images, without explaining their main differences and similarities. Our metric, on the other hand, provides visual explanations to support the calculated score, making the similarity assessment transparent and justified.

Figures (8)

• Figure 1: Comparison of human perceptual similarity and nonperceptual metrics. On a sample of the 2AFC dataset zhang_unreasonable_2018, the similarity metrics PSNR and SSIM wang_mean_2009 assigned the highest similarity score between the reference image and the distorted one on the right, suggesting that, for both metrics, the reference image is more similar to the distorted image on the right than the distorted image on the left. However, a panel of five human judges unanimously agreed that based on their visual perception, the reference image is more similar to the left distorted image.
• Figure 2: The following example illustrates the process of creating a signature from an image. The image is converted to grayscale and split into patches. Subsequently, Gabor's filter dictionary is applied to each patch. The energy magnitudes are calculated for each Gabor response, resulting in a 4x6 matrix for each patch. To reduce the number of patches, the MengHee-Heng clustering algorithm is applied to the matrices, yielding k 24-dimensional centroids V and the corresponding frequencies W. V and W are then output to the S dictionary.
• Figure 3: Displacement of all $256^3$ possible RGB color combinations in the Oklab perceptual color space.
• Figure 4: Correspondences of similarity metrics with the 2AFC dataset. The horizontal axis represents the perceptual similarity of humans who labeled the 2AFC dataset, as reported in zhang_unreasonable_2018. The results show that, compared with other low-level FR methods, EDOKS is the closest similarity metric to human perceptual similarity, while it is close to the results of other deep metrics.
• Figure 5: Average scores responses obtained by EDOKS and the SOTA metrics calculated on the JND dataset at the pairs of images rated same (green) and not same (red) by humans. The discrepancy in scores between same and not same subsets obtained by our EDOKS metric shows behavior more consistent with human perception than the scores obtained by other low level metrics, while it is close to the results of other deep metrics.