[Paper Review] Quantitative Evaluation of Base and Detail Decomposition Filters Based on their Artifacts

Charles Hessel, Jean‐Michel Morel|arXiv (Cornell University)|Aug 28, 2018

Image Enhancement Techniques44 references3 citations

TL;DR

This paper proposes a quantitative framework to objectively evaluate base and detail decomposition filters by identifying and measuring four key artifacts—luminance halo, contrast halo, staircasing, and compartmentalization—using custom test patterns and mathematical formulas. It introduces a cross-calibrated, expert-annotated fusion score that ranks filters based on artifact severity, with the fast local Laplacian filter (FLL) emerging as the top-performing method among seven tested filters.

ABSTRACT

This paper introduces a quantitative evaluation of filters that seek to separate an image into its large-scale variations, the base layer, and its fine-scale variations, the detail layer. Such methods have proliferated with the development of HDR imaging and the proposition of many new tone-mapping operators. We argue that an objective quality measurement for all methods can be based on their artifacts. To this aim, the four main recurrent artifacts are described and mathematically characterized. Among them two are classic, the luminance halo and the staircase effect, but we show the relevance of two more, the contrast halo and the compartmentalization effect. For each of these artifacts we design a test-pattern and its attached measurement formula. Then we fuse these measurements into a single quality mark, and obtain in that way a ranking method valid for all filters performing a base+detail decomposition. This synthetic ranking is applied to seven filters representative of the literature and shown to agree with expert artifact rejection criteria.

Motivation & Objective

• To address the lack of objective, standardized evaluation for base and detail decomposition filters used in tone mapping and contrast enhancement.
• To identify and mathematically characterize the most prevalent and perceptually significant artifacts in such filters.
• To develop a fair comparison methodology by cross-calibrating filter parameters to ensure consistent enhancement levels across methods.
• To fuse independent artifact measurements into a single, interpretable quality score using expert rejection criteria.
• To rank existing filters objectively and validate the ranking against expert judgments.

Proposed method

• Define four primary artifacts: luminance halo, contrast halo, staircasing (contrast reversal), and compartmentalization, each with a distinct mathematical characterization.
• Design specific test patterns for each artifact to provoke and isolate its presence in any filter output.
• Develop a measurement formula for each artifact based on the filter's response to its corresponding test pattern.
• Implement a cross-calibration procedure to align filter parameters so that all filters produce the same level of detail enhancement, enabling fair comparison.
• Use expert annotations to determine acceptability thresholds for each artifact, then normalize artifact scores using these thresholds.
• Fuse the normalized artifact scores into a single quality mark using a weighted sum of squared normalized scores: $\mathcal{A}(f) = \frac{\mathcal{H}(f)^2}{\mathcal{H}(\text{IS-}L^0)^2} + \frac{\mathcal{S}(f)^2}{\mathcal{S}(\text{FBF})^2} + \frac{\mathcal{P}(f)^2}{\mathcal{P}(\text{WLS})^2} + \frac{\mathcal{C}(f)^2}{\mathcal{C}(\text{GF})^2}$.

Experimental results

Research questions

• RQ1 Which artifacts most significantly degrade the perceptual quality of base and detail decomposition filters?
• RQ2 How can each artifact be mathematically defined and isolated using controlled test patterns?
• RQ3 What is a fair and consistent method to calibrate different filters so they can be compared under equivalent enhancement conditions?
• RQ4 How can independent artifact measurements be fused into a single, meaningful quality score that reflects expert rejection behavior?
• RQ5 Which base and detail decomposition filter performs best under this objective, artifact-based evaluation?

Key findings

• The fast local Laplacian filter (FLL) achieved the lowest overall artifact score and was ranked first, indicating superior performance in minimizing all four artifacts.
• The image smoothing via $L^0$ gradient minimization (IS-$L^0$) filter was invalidated by the luminance halo and staircasing artifacts, ranking among the worst.
• The domain transform (DT) filter failed due to luminance halo and compartmentalization, and was ranked lowest.
• The guided filter (GF) was rejected primarily due to the contrast halo, and ranked poorly despite strong performance on other metrics.
• Filters such as FBF and WLS were invalidated by two or more artifacts, confirming their limited suitability for high-quality contrast enhancement.
• The final ranking based on the fused score $\mathcal{A}(f)$ strongly agreed with expert rejection decisions, validating the method’s reliability.