Edge Preservation Index Explained

• Edge Preservation Index is defined by EPRa and EPRr metrics that compare binary Canny edge maps from reference and processed images.
• It captures both accuracy and robustness of edge fidelity, complementing traditional metrics like PSNR, SSIM, and others in evaluating interpolation methods.
• Alternative approaches include gradient-space losses for denoising and structural analysis of epipolar plane images in light-field reconstruction.

Searching arXiv for the cited papers and closely related terminology. In the arXiv literature considered here, Edge Preservation Index is not a canonical paper title term but a useful umbrella description for quantitative measures of how well image-processing or reconstruction outputs retain the edge structure of a reference image. The most explicit formulation appears in "Performance Evaluation of Edge-Directed Interpolation Methods for Images" (Yu et al., 2013), which introduces two edge-preserving ratios—EPRa and EPRr—as scalar measures of edge-preserving ability. Related work on denoising introduces an edge preservation loss based on Sobel gradients rather than a named index (Ofir et al., 2018), while light-field reconstruction uses the acronym EPI in a different sense, namely epipolar plane image, and evaluates edge and structure preservation indirectly through EPI line geometry, PSNR, and MS-SSIM rather than an explicit edge-preservation index (Wu et al., 2021).

1. Terminological scope and usage

The most precise arXiv source for an edge-preservation metric is (Yu et al., 2013). That paper states that “no parameters are mentioned to measure edge-preserving ability of edge-adaptive interpolation approaches and we proposed two,” and then defines EPRa and EPRr as the paper’s dedicated edge-preservation measures (Yu et al., 2013). Although the paper does not explicitly use the name Edge Preservation Index or the acronym EPI, the two ratios are scalar quantities in $[0,1]$ intended to quantify edge preservation, and therefore function as edge preservation indices in the ordinary methodological sense (Yu et al., 2013).

A recurrent source of ambiguity is that EPI already has an established meaning in light-field imaging: epipolar plane image. In "Light Field Reconstruction Using Convolutional Network on EPI and Extended Applications" (Wu et al., 2021), EPI denotes a 2D slice $E_{y^*,t^*}(x,s)$ of a 4D light field $L(x,y,s,t)$, not an edge-preservation score (Wu et al., 2021). This distinction is important because that paper is centrally concerned with preserving line and edge structures in EPIs, yet it does not define an Edge Preservation Index.

A second terminological extension arises in denoising. "Multi-scale Processing of Noisy Images using Edge Preservation Losses" (Ofir et al., 2018) does not define a named edge-preservation index either; instead it introduces an edge preservation loss that compares Sobel-derived gradients of clean and denoised images (Ofir et al., 2018). This suggests a broader usage in which edge preservation can be operationalized either as an evaluation index, as in (Yu et al., 2013), or as a training objective, as in (Ofir et al., 2018).

2. Formal definitions in edge-directed interpolation

Section 2 of (Yu et al., 2013), “Edge-Preserving Ability,” provides the explicit definitions. The paper states that Canny edge maps are extracted from the standard image and from the interpolated image, denoted $EM_s$ and $EM_i$, and then defines:

$EPR_a = \frac{|EM_s \cap EM_i|}{|EM_s|}$

and

$EPR_r = \frac{|EM_s \cap EM_i|}{|EM_s \cup EM_i|}$

with the set operations corresponding to intersection and union of edge pixels (Yu et al., 2013).

The paper interprets EPRa as the edge-preserving ratio “from accuracy” and EPRr as the edge-preserving ratio “from robustness” (Yu et al., 2013). In implementable form, $EM_s$ is the set of edge pixels in the standard image and $EM_i$ is the set of edge pixels in the interpolated image, both produced by the Canny detector with automatically settled thresholds (Yu et al., 2013). The formulas must be read as ratios of set cardinalities rather than literal set division.

These two measures capture distinct aspects of edge preservation. EPRa measures how many reference edge pixels remain edges after interpolation, so it behaves as an edge recall measure and penalizes missed edges without penalizing spurious extra edges (Yu et al., 2013). EPRr measures the overlap relative to the union of detected edges and therefore penalizes both missing and spurious edges; (Yu et al., 2013) explicitly frames it as the parameter “from robustness.” The paper’s structured explanation further notes that this is mathematically analogous to the Jaccard index between the two edge sets (Yu et al., 2013).

The range and interpretation are direct. For both metrics, values lie in $[0,1]$, larger values are better, and the upper bound corresponds to perfect agreement of preserved edges under the paper’s binary edge-map definition (Yu et al., 2013).

3. Computation protocol and experimental context

The computation protocol in (Yu et al., 2013) is tightly specified. A high-resolution standard image $E_{y^*,t^*}(x,s)$0 is downsampled according to

$E_{y^*,t^*}(x,s)$1

which keeps the top-left pixel in each $E_{y^*,t^*}(x,s)$2 block (Yu et al., 2013). Interpolation methods are then applied to reconstruct a high-resolution image $E_{y^*,t^*}(x,s)$3, and Canny edge maps are extracted from both $E_{y^*,t^*}(x,s)$4 and $E_{y^*,t^*}(x,s)$5 using automatic thresholding (Yu et al., 2013).

The resulting workflow is: reference high-resolution image, deterministic downsampling, interpolation back to the original size, Canny edge extraction, and then computation of $E_{y^*,t^*}(x,s)$6 and $E_{y^*,t^*}(x,s)$7 from the binary edge maps (Yu et al., 2013). No edge directions, edge strengths, or gradient magnitudes enter the metric definitions themselves; the measures operate only on binary edge presence or absence (Yu et al., 2013).

The experimental setting includes two image groups: 12 digital images of size $E_{y^*,t^*}(x,s)$8, and 12 fetal spine MR images of size $E_{y^*,t^*}(x,s)$9 (Yu et al., 2013). Six interpolation methods are compared: bi-linear, bi-cubic, NEDI, EGII, ICBI, and DCCI (Yu et al., 2013). The edge-preservation measures are evaluated alongside SNR, PSNR, SSIM, FSIM, Mutual Information, and Time Cost (Yu et al., 2013).

The reported averages make the role of the edge-preservation measures concrete. For digital images, ICBI attains the highest $L(x,y,s,t)$0 and $L(x,y,s,t)$1, followed closely by DCCI and EGII, while bi-linear and bi-cubic obtain markedly lower values (Yu et al., 2013). For MR images, ICBI again attains the highest $L(x,y,s,t)$2 and $L(x,y,s,t)$3, with DCCI and EGII next, and traditional methods again lowest (Yu et al., 2013). The paper concludes from Tables 1 and 2 that EDI methods are better than traditional methods except time cost, and that higher EPRa and EPRr accompany higher SNR, PSNR, SSIM, FSIM, and MI (Yu et al., 2013).

4. Interpretation, significance, and limitations

The main significance of EPRa and EPRr is that they evaluate edge preservation directly rather than inferring it from generic image-quality measures. (Yu et al., 2013) presents them precisely because existing evaluation practice did not supply a dedicated parameter for edge-preserving ability in edge-adaptive interpolation (Yu et al., 2013). They therefore complement PSNR- or SSIM-style fidelity metrics with a structural criterion targeted at the design objective of edge-directed interpolation.

The distinction between the two measures is methodologically important. EPRa emphasizes preservation of true edges present in the reference image; EPRr additionally penalizes edge hallucinations and thus acts as a stricter measure of structural consistency (Yu et al., 2013). A plausible implication is that the pair $L(x,y,s,t)$4 is more informative than either metric alone when evaluating methods that may trade off smoothing against false-edge generation.

The paper also records a nuance that can be misread if only the scalar scores are considered. It notes that “from edge-generation aspect, traditional interpolation methods generate more details than that of EDI methods for their EPR parameters,” even though traditional methods have lower EPR scores (Yu et al., 2013). In context, this indicates that edge generation per se is not equivalent to faithful edge preservation: additional detected edges may reflect spurious structure and reduce robustness rather than improve it.

Several limitations are explicit or directly implied in (Yu et al., 2013). The measures depend entirely on the chosen edge detector, which is Canny with automatic threshold selection in that study (Yu et al., 2013). They are binary and therefore do not measure edge strength, orientation accuracy, or sub-pixel displacement (Yu et al., 2013). They are also sensitive to small geometric misalignments, since a one-pixel shift can reduce intersection and enlarge union even when the visual difference is modest (Yu et al., 2013). These are not incidental details; they determine what kind of structural fidelity the indices do and do not capture.

5. Gradient-space formulations in denoising

"Multi-scale Processing of Noisy Images using Edge Preservation Losses" (Ofir et al., 2018) broadens the conceptual landscape by defining edge preservation in gradient space rather than via binary edge-set overlap. In Section 5, the paper introduces an auxiliary loss for denoising:

$L(x,y,s,t)$5

where $L(x,y,s,t)$6 is the clean image, $L(x,y,s,t)$7 is the noisy input, and $L(x,y,s,t)$8 is the denoising network (Ofir et al., 2018). The text describes the label as the Sobel edge detector applied to the clean image and compared to the edge maps of the denoised image (Ofir et al., 2018).

This formulation does not yield a named edge-preservation index in the paper’s tables; it is used as a training loss, added to the reconstruction loss:

$L(x,y,s,t)$9

(Ofir et al., 2018). Nonetheless, it operationalizes edge preservation as similarity between the clean and processed image in gradient space. The paper states that this edge preserving auxiliary loss improves the performance of IDCNN in denoising in both PSNR and SSIM (Ofir et al., 2018).

The contrast with (Yu et al., 2013) is instructive. In (Yu et al., 2013), preservation is evaluated through binary Canny edge maps and overlap ratios. In (Ofir et al., 2018), preservation is enforced through squared $EM_s$0 difference of Sobel-derived gradients along $EM_s$1 (Ofir et al., 2018). This suggests two distinct but compatible paradigms for edge preservation: set-overlap measures for evaluation and gradient-discrepancy measures for optimization.

The paper also notes an important limitation of its own explicit formulation: only the horizontal derivative is written in the loss (Ofir et al., 2018). The structured explanation accompanying the paper indicates that extending the method to full gradient magnitude would be consistent with the core idea, but such an extension is not the paper’s explicit loss definition (Ofir et al., 2018). For an encyclopedia treatment, this distinction matters because it separates what is directly stated from what is merely a natural generalization.

6. Distinguishing edge-preservation indices from epipolar plane images

The acronym EPI has a different and well-established meaning in light-field reconstruction. In (Wu et al., 2021), an epipolar plane image is a 2D slice $EM_s$2 obtained from the 4D light field $EM_s$3 by fixing one spatial coordinate and one angular coordinate (Wu et al., 2021). Scene points become approximately straight lines in the EPI, with line slopes encoding disparity and hence depth (Wu et al., 2021).

That paper is highly relevant to structural preservation because it is built around maintaining EPI line continuity and avoiding ghosting under angular undersampling. It introduces a blur–restoration–deblur framework for reconstructing angular details in EPIs and evaluates results with PSNR, MS-SSIM, error maps, and visual analysis of EPI line structures (Wu et al., 2021). However, it does not define an explicit metric called Edge Preservation Index (Wu et al., 2021).

This distinction addresses a common misconception. In light-field literature, “EPI preservation” ordinarily means preservation of epipolar plane image structure, not computation of an edge-preservation score. The paper’s concern with ghosting, aliasing, line continuity, and depth enhancement shows that EPI geometry is a structural fidelity object, but the evaluation remains indirect: high MS-SSIM, sharp EPI lines, and improved depth RMSE act as evidence of good structure preservation rather than as a named edge-preservation index (Wu et al., 2021).

A plausible implication is that the phrase Edge Preservation Index should not be abbreviated as EPI in contexts where epipolar plane images are already under discussion, because the acronym collision is substantive rather than merely stylistic.

7. Synthesis and methodological position

Across these arXiv sources, edge preservation appears in three analytically distinct forms. First, (Yu et al., 2013) provides explicit edge-preservation ratios $EM_s$4 and $EM_s$5, which are the clearest examples of edge preservation indices in the strict evaluative sense. Second, (Ofir et al., 2018) uses an edge preservation loss in gradient space to train denoising networks, thereby treating edge preservation as an optimization target rather than a standalone benchmark metric. Third, (Wu et al., 2021) focuses on preservation of EPI line structure in light fields, where edge fidelity is central but not formalized through a dedicated edge index.

These formulations are complementary rather than contradictory. Binary edge-set overlap, gradient-space discrepancy, and structural continuity in epipolar plane images all target the same underlying concern: whether processing preserves salient discontinuities and the geometric information they encode. What changes is the representation of “edge”—binary Canny maps in (Yu et al., 2013), Sobel gradients in (Ofir et al., 2018), and line geometry in EPIs in (Wu et al., 2021).

For image interpolation, the literature in (Yu et al., 2013) supports the strongest concrete definition: an edge preservation index can be understood as a scalar ratio comparing edge maps of reference and output, with separate emphasis on accuracy and robustness. For denoising, (Ofir et al., 2018) suggests that gradient fidelity can serve as a practical surrogate for edge preservation. For light-field reconstruction, (Wu et al., 2021) shows that edge preservation extends beyond 2D image boundaries into multi-view structural consistency, even when the relevant acronym EPI denotes something else entirely.

Under this synthesis, the most defensible encyclopedia definition is narrow but explicit: in the available arXiv evidence, an Edge Preservation Index is best represented by the pair of edge-preserving ratios proposed in (Yu et al., 2013), namely $EM_s$6 and $EM_s$7, while related literature uses edge-preservation losses or structural surrogates when the exact term is absent (Yu et al., 2013).