{"id":"0ec9f5b6-e37a-4726-ba42-4e58ff45c676","arxiv_id":"1908.06307","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A multi-kernel bilateral filter that derives its range-kernel widths from a hierarchical clustering of the image, reported to beat BF, TV, and CF on nonstationary Gaussian noise.","lead":"This paper proposes a denoising filter that changes how strongly it smooths different parts of an image based on the image's own structure, using a clustering tree to set the filter's local strength. The authors claim it outperforms three classical filters on natural photos and synthetic MRI data with uneven noise, though the comparison omits modern denoising methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 11's contextual factor acts opposite to the prose: tau<1 shrinks the range kernel and tau->0 blocks cross-boundary smoothing, so the mechanism as written cannot deliver the claimed adaptive behavior.","rationale":"The reader's verdict is REJECT and my analysis supports that outcome, but for a slightly different primary reason. The reader's weakest_assumption is that leaf sigma_t,k estimated from the noisy image reflects noise rather than image structure; that is a genuine identifiability concern about how the range width is chosen. The more decisive problem is internal: Eq. 11 and Eq. 12 define a kernel whose dependence on tau is the opposite of the behavior described in Section 3.2.2. This is not about missing baselines or error bars; if the formula is taken literally, the contextual mechanism cannot do what the paper says it does. The empirical comparison is thin (only BF, TV, and CF; no code or data; no error bars), which weakens support for the headline claim, but the equation-prose conflict is more load-bearing because it attacks the construction itself. A concrete numerical check with a two-level tree will settle the direction of the effect. I therefore keep the reader's REJECT verdict. Agreement is partial because the reader's rationale flagged this inconsistency even though the formal weakest_assumption was about sigma_t,k; both concerns point in the same direction, but the Eq. 11 conflict is the more fundamental one.","tokens_in":946,"tokens_out":1031,"duration_ms":95032,"concrete_test":"Take a two-level cluster tree with sigma_{t-2}=30, sigma_{t-1}=10, sigma_t=5, so tau=1/3, and compute the Eq. 11 range weight for an intensity difference Delta=10: exp(-100/(2*(1/3)*25)) = exp(-6) ~ 0.0025, much smaller than the weight exp(-100/(2*25)) ~ 0.135 that the prose's 'enlarged' kernel would imply. Also set tau=0.1 and observe that the weight tends to 0, whereas Section 3.2.2 says tau close to 0 enlarges sigma and smooths across C_{t,k}. If a corrected equation places tau in the numerator, rerun the same numerical check and verify that the contextual behavior matches the prose; this single comparison will show whether Eq. 11 as written can implement the claimed mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is the contextual factor tau_{t,k} in Eq. 11. As written, the range kernel is exp( -(I(x)-I(xi))^2 / (2 tau_{t,k} sigma_{t,k}^2) ), with tau_{t,k}=sigma_{t-1,k*}/sigma_{t-2,k**} (Eq. 12). Section 3.2.2 states that sigma_{t-2}>sigma_{t-1}, hence tau<1, which 'enlarges sigma_{t,k}', and that tau close to 0 'significantly enlarges sigma' and permits smoothing across the boundaries of leaf cluster C_{t,k}. Equation 11 does the opposite: for tau<1 the effective range width is sqrt(tau) sigma_{t,k}, which is narrower than the leaf variance, and as tau->0 the kernel becomes a delta that preserves every intensity difference, preventing any cross-boundary smoothing. The claimed adaptive behavior is therefore not the behavior of the stated formula. Since this contextual factor is the only place where image context enters the filter, the central claim rests on an internally inconsistent specification. A typographical correction might repair it, but the manuscript as submitted does not allow a reader to verify which behavior was intended.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Multi-Kernel Filter (MKF), an extension of bilateral filtering designed for nonstationary noise. It constructs a hierarchical cluster tree of image pixels via EM-based similarity clustering and proximity clustering, then uses leaf-cluster standard deviations and ratios of predecessor-cluster standard deviations to set spatially varying range-kernel widths. The authors claim that MKF adapts filtering kernels to image content automatically and report experiments on BSD300 and BrainWeb showing lower MAE and higher SSIM than BF, TV, and CF.","tokens_in":10772,"tokens_out":7165,"duration_ms":60047,"significance":"If the mechanism were correctly specified, the idea of using a cluster-tree context to set per-pixel range kernels would be a plausible contribution to adaptive image filtering, and the qualitative kernel illustrations in Figure 5 are suggestive. However, the central formula in Eq. (11) contradicts the accompanying prose, and the experimental evaluation is too thin to support the 'state-of-the-art' claim. The method's adaptivity is shown in Figure 4 to depend strongly on a manually chosen tree depth, and the paper does not provide code, error bars, or statistical tests. The idea is promising but the current manuscript does not establish its validity.","major_comments":[{"comment":"The contextual factor as written does the opposite of what the text claims. In Eq. (11), the range-kernel denominator is 2 τ_{t,k} σ_{t,k}^2; since the paper states that σ_{t−2,k**} > σ_{t−1,k*}, Eq. (12) yields τ_{t,k} < 1, so the effective range variance τ_{t,k} σ_{t,k}^2 is smaller than σ_{t,k}^2. As τ_{t,k} → 0, the kernel approaches a delta in intensity, which blocks smoothing across intensity differences. The text in §3.2.2 instead states that τ close to 0 'significantly enlarges σ_{t,k}' and 'smooths across the boundaries' of C_{t,k}. This is an internal contradiction in the central mechanism, and the reader cannot determine whether Eq. (11) or the prose describes the intended behavior.","section":"§3.2.2, Eq. (11)-(12)"},{"comment":"Figure 4 shows that the optimal tree depth Dp depends strongly on the noise level: for noise level 10, t=7 gives the best MAE/SSIM, while for noise level 1000, t=7 gives the worst. Since Dp is a manually initialized parameter (§4.1), the method does not automatically adapt to noise nonstationarity; a user must know the noise level to choose Dp. This contradicts the central claim of automatic adaptivity and is acknowledged in the conclusion ('its performance depended on the parameters of the clustering'). The near-linear curves for t=2 in Fig. 6 are achieved by fixing Dp=2, not by adaptation.","section":"§4.3, Fig. 4"},{"comment":"The experimental evidence is too limited to support the claim that MKF 'outperforms state-of-the-art filters.' On BSD300, MKF is compared only with BF; on BrainWeb, it is compared with BF, TV, and CF, all from 1998, 1992, and 2017 respectively, which are not state-of-the-art denoisers for the reported tasks. No error bars, standard deviations, or statistical significance tests are reported. Also, no sensitivity analysis is given for baseline parameters beyond the statement that they were 'well chosen.' These omissions leave the quantitative superiority claim unsubstantiated.","section":"§4.2-4.4"},{"comment":"The adaptivity of MKF rests on the assumption that the leaf-cluster standard deviation σ_{t,k}, estimated by EM from the noisy image (Eqs. (2)-(3)), reflects local noise statistics rather than image structure. In natural images, σ_{t,k} is a mixture of both, and using it directly as the range-kernel width may over-smooth textured regions or under-smooth noisy flat regions. No experiment validates this assumption, for example by comparing the estimated σ_{t,k} with the ground-truth noise level in the synthetic BrainWeb data.","section":"§3.1.1-§3.2"},{"comment":"The definition of τ_{t,k} is not reproducible for t=2. Eq. (12) references σ_{t−2,k**}, which does not exist for t=2, but §4.3 states that τ_{t,k} = (σ_{t,k}/σ_{t−1,k*})^2 when t=2. These two definitions disagree, and the manuscript does not explain how the root-layer case is handled. This ambiguity affects the parameter setting used in the main experiments.","section":"§4.3, Eq. (12)"}],"minor_comments":[{"comment":"The affiliation line contains a typo: 'Universtiy' should be 'University'.","section":"§1 affiliation"},{"comment":"The parentheses are unbalanced in 'O(I(x)))'; one closing parenthesis should be removed.","section":"§3.2.1, Eq. (4)"},{"comment":"The cross-reference 'In Figure 4.3' should read 'In Figure 4'.","section":"§4.3, second paragraph"},{"comment":"The abstract duplicates a sentence: 'We first design...' appears twice, and the 'At first, we design...' sentence repeats the same idea verbatim.","section":"Abstract"},{"comment":"The name 'Conduction coefficient' for the spatial kernel variance h_x is nonstandard; consider using 'spatial kernel width' instead.","section":"§4.1, parameter list"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read the MKF paper. The core idea is genuine: replace BF's single range kernel with multiple kernels derived from a hierarchical cluster tree. Leaf cluster variance sets the range width per pixel, and ancestor clusters provide a contextual factor. That combination is new relative to the cited literature, and the clustering procedure—EM on intensity followed by proximity grouping—is described clearly enough to reproduce. The authors also frame a real problem: nonstationary noise in MRI and natural images.\n\nThe paper is not ready for publication, though, because of a load-bearing inconsistency. Eq. 11 puts tau_{t,k} in the denominator of the exponent, so the range Gaussian has variance tau * sigma^2. The prose in Section 3.2.2 says tau<1 \"enlarges\" sigma and that tau close to 0 allows smoothing across cluster boundaries. But tau<1 narrows the kernel, and tau->0 makes it a delta, preserving every intensity difference. So the mechanism as written does the opposite of what the text claims. The contextual factor is the only place where image context enters the filter, so this contradiction cuts to the heart of the method. If it's a typo, a corrected equation might make the method behave as intended; but a reader cannot verify which behavior was actually tested.\n\nThe experimental support is also thin. Only three baselines—BF, TV, CF—no error bars, no statistical tests, no code or data. Calling those \"state-of-the-art\" is generous. Figure 4 shows tree depth Dp must be tuned to noise level; the authors admit the performance depends on clustering parameters. That undercuts the \"automatic\" adaptivity claim in the abstract.\n\nWhat's good: the idea is worth developing. The multi-scale cluster representation is a sensible way to make BF context-aware. The MRI experiments on complex data are a nice touch. But the current manuscript doesn't provide enough evidence to support the claims.\n\nMy recommendation: this deserves peer review, because the core idea is novel and the main flaw might be fixable. A competent referee could catch the equation inversion and demand a corrected version with proper baselines. I wouldn't accept it in this form, but I wouldn't bury it either.\n\nFor you: if you work on adaptive filtering, it's worth a skim. The cluster-tree construction might be reusable even if the noise model isn't.","headline":"A genuinely new way to make bilateral filtering context-aware, but the central equation contradicts its own prose and the experiments are too thin to support the claims.","tokens_in":11334,"tokens_out":3028,"would_cite":false,"duration_ms":29468,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing the single fixed range kernel of bilateral filtering with cluster-tree-derived kernels lets it handle nonstationary noise automatically.","keywords":["bilateral filtering","multi-kernel filter","nonstationary noise","image denoising","hierarchical clustering","spatially varying noise","adaptive filtering","structural similarity"],"falsifier":"Take a clean image containing a known flat region and a sharp edge, add Gaussian noise of known standard deviation, run MKF's clustering, and record the leaf-cluster variances $\\sigma_{t,k}$ for flat-region and edge-region clusters. If the estimated variance rises with edge contrast rather than matching the injected noise level, the kernel widths are tracking image structure, not noise, and the adaptivity claim fails.","tokens_in":10308,"feed_emoji":"🖼️","tokens_out":12429,"duration_ms":116456,"temperature":0.7,"pith_summary":"Bilateral filtering is a widely used denoising method, but its range kernel width is a free parameter that presumes one global noise level, so variable noise defeats it. The paper proposes a multi-kernel filter (MKF) that learns the kernels from the image itself: a hierarchical clustering algorithm partitions the noisy image into a cluster tree, and each leaf cluster, together with two ancestor clusters, sets a local range-kernel width at every pixel. The authors report that one fixed MKF configuration maintains near-linear error curves across noise levels 10 to 1000 on BSD300 and that MKF also handles spatially varying noise on BrainWeb MRI, outperforming BF, TV, and CF in both MAE and SSIM. The practical importance is that the filter's most sensitive parameter, the noise scale, no longer needs manual initialization per image.","feed_headline":"Self-tuning kernels beat fixed bilateral filtering on noisy images","feed_subtitle":"On two public datasets, automatic kernels remove the manual tuning that breaks bilateral filtering under variable noise.","key_machinery":"The load-bearing object is the cluster tree: a coarse-to-fine hierarchy of coherent pixel groups built by iterative two-class EM on intensity (similarity clustering) followed by connected-component isolation (proximity clustering). It supplies the image context. Each leaf cluster $C_{t,k}$ contributes one range kernel for the pixels it contains, and its two ancestor clusters $C_{t-1,k^*}$ and $C_{t-2,k^{**}}$ contribute the contextual scaling factor $\\tau_{t,k} = \\sqrt{\\sigma_{t-1,k^*} \\sigma_{t-2,k^{**}}}$. The kernel in Eq. (11), $w(\\vec x, \\vec \\xi \\mid \\sigma_{t,k}, \\tau_{t,k}) = \\exp\\big[-(\\vec x - \\vec \\xi)^2 / (2 h_{\\vec x}^2) - (I(\\vec x)-I(\\vec \\xi))^2 / (2 \\tau_{t,k} \\sigma_{t,k}^2)\\big]$, sets the effective range width pixel by pixel. Thus the same local intensity gradient is smoothed heavily inside low-salience regions and preserved near salient boundaries; this is what makes the single fixed kernel of BF into multiple automatically learned, spatially varying kernels.","core_discovery":"Bilateral filtering rests on one manually chosen range-kernel width $h_I$, which encodes a global noise level; when the noise level changes over the image or across images, that single width can be either too small to smooth noisy pixels or too large to preserve edges. The paper's core claim is that this single kernel should be replaced by many kernels whose widths are learned from the image itself. MKF builds a cluster tree from the noisy input, with each leaf cluster characterized by an estimated mean $\\mu_{t,k}$ and standard deviation $\\sigma_{t,k}$, and with the variances of two ancestor clusters entering through $\\tau_{t,k} = \\sqrt{\\sigma_{t-1,k^*} \\sigma_{t-2,k^{**}}}$. The effective range width at a pixel is then $\\tau_{t,k} \\sigma_{t,k}$. With this construction, the paper reports that on BSD300 with noise levels 10 to 1000, and on BrainWeb MRI with spatially varying Gaussian noise, MKF outperforms BF, TV, and CF on both MAE and SSIM.","pith_inferences":["Editorial inference: The cluster-tree conditioning is a general template; replacing the intensity-variance statistic with edge orientation or texture energy would give context-conditioned kernels for other bilateral-filter pipelines, such as tone mapping, depth upsampling, or edge-preserving smoothing.","Editorial inference: A direct attribution experiment is available: compare MKF against BF whose range width is set from the ground-truth noise-level map. If oracle BF matches MKF, the benefit reduces to per-pixel noise estimation; if it does not, the two-ancestor context term adds more than variance information.","Editorial inference: The Gestalt and visual-context narration is not required by the mathematics; the same equations can be read as a hierarchical mixture estimate of local noise scale, so the method can be adopted without committing to that framing."],"forward_implications":["With one fixed configuration, MKF produces near-linear MAE and SSIM curves across noise levels 10 to 1000 on BSD300, whereas BF's curves bend sharply and depend on the hand-set kernel width.","On complex-valued BrainWeb MRI corrupted by spatially varying Gaussian noise, MKF yields lower MAE and higher SSIM than BF, TV, and CF, and its output keeps the phase-induced gradual transitions rather than producing mosaic blocks.","Because the clustering splits the image recursively with fixed rules, the method adapts the number of clusters to each image automatically, removing the manual membership count that flat clustering requires.","The two-ancestor context term makes the filter smooth across boundaries of low-salience leaf clusters while preserving boundaries of high-salience ones, which is the mechanism behind the reported texture preservation."],"supporting_citations":[{"why":"Defines bilateral filtering; MKF is explicitly an extension of its single range kernel.","marker":"Tomasi and Manduchi, 1998"},{"why":"Reinterprets BF in a generalized intensity space, the formulation MKF modifies.","marker":"Barash, 2002"},{"why":"Total variation filter used as a comparison baseline on spatially varying noise.","marker":"Rudin, Osher and Fatemi, 1992"},{"why":"Curvature filter baseline, the second comparison method on BrainWeb.","marker":"Gong and Sbalzarini, 2017"},{"why":"Supplies the BSD300 dataset used for the integrally varying noise experiments.","marker":"Martin, Fowlkes, Tal and Malik, 2001"},{"why":"Supplies the BrainWeb T1 volume used as ground truth for spatially varying MRI noise experiments.","marker":"Kwan, Evans and Pike, 1999"},{"why":"Provides the phase-correction setup used to synthesize complex MRI components with spatially varying noise.","marker":"Pizzolato et al., 2016"},{"why":"Justifies the spatially varying Gaussian noise model assumed for the MRI experiments.","marker":"Aja-Fernández, Pie, Vegas-Sánchez-Ferrero et al., 2015"}],"fun_headline_variants":["MKF adapts filters per image, beating fixed kernels on variable noise","Cluster tree guides kernel widths to beat bilateral filtering under changing noise","Automatic multi-kernel filter outperforms fixed bilateral on nonstationary noise","Image context tunes each pixel's filter kernel for better denoising","From one kernel to many: adaptive filtering wins on variable-noise images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the intensity spread estimated from each small cluster of the noisy image measures local noise level rather than the scene's edges or texture; if the clustering encodes structure, then every adapted kernel width is off, and the filter's adaptivity no longer means what the paper claims.","fun_headline_variants_meta":{"raw":{"variants":["MKF adapts filters per image, beating fixed kernels on variable noise","Cluster tree guides kernel widths to beat bilateral filtering under changing noise","Automatic multi-kernel filter outperforms fixed bilateral on nonstationary noise","Image context tunes each pixel's filter kernel for better denoising","From one kernel to many: adaptive filtering wins on variable-noise images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000844,"raw_usage":{"total_tokens":3738,"prompt_tokens":1068,"completion_tokens":2670,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":2577}},"tokens_in":684,"tokens_out":2670,"duration_ms":17510,"temperature":1.0,"reasoning_tokens":2577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:49:46.323341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a clean image containing a known flat region and a sharp edge, add Gaussian noise of known standard deviation, run MKF's clustering, and record the leaf-cluster variances $\\sigma_{t,k}$ for flat-region and edge-region clusters. If the estimated variance rises with edge contrast rather than matching the injected noise level, the kernel widths are tracking image structure, not noise, and the adaptivity claim fails.","supporting_citations":[{"cited_title":"Bilateral filtering for gray and color images, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), pp","cited_arxiv_id":null,"evidence_quote":"Defines bilateral filtering; MKF is explicitly an extension of its single range kernel."},{"cited_title":"Fundamental relationship between bilateral filtering, adaptive smoothing, and the nonlinear diffusion equation","cited_arxiv_id":null,"evidence_quote":"Reinterprets BF in a generalized intensity space, the formulation MKF modifies."},{"cited_title":"Nonlinear total variation based noise removal algorithms.Physica D: Nonlinear Phenomena 60, 259–268","cited_arxiv_id":null,"evidence_quote":"Total variation filter used as a comparison baseline on spatially varying noise."},{"cited_title":"Curvature filters efficiently reduce certain variational energies","cited_arxiv_id":null,"evidence_quote":"Curvature filter baseline, the second comparison method on BrainWeb."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BSD300 dataset used for the integrally varying noise experiments."},{"cited_title":"MRI simulation-based evaluation of image-processing and classification methods","cited_arxiv_id":null,"evidence_quote":"Supplies the BrainWeb T1 volume used as ground truth for spatially varying MRI noise experiments."},{"cited_title":"Noise floor removal via phase correction of complex diffusion-weighted images: Influence on DTI and Q-space metrics, in: MICCAI Workshop on Computational Diffusion MRI, pp","cited_arxiv_id":null,"evidence_quote":"Provides the phase-correction setup used to synthesize complex MRI components with spatially varying noise."},{"cited_title":"Spatially variant noise estimation in MRI: A homomorphic approach","cited_arxiv_id":null,"evidence_quote":"Justifies the spatially varying Gaussian noise model assumed for the MRI experiments."}],"review_version":1}