{"id":"35408547-3865-454d-9b01-ef51e09d5c30","arxiv_id":"2505.02365","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"QMCIF decomposes quaternion-represented color images into base, detail, and noise layers, builds dual-scale focus maps, and selects the sharpest patches to fuse multi-focus color images into one sharp output.","lead":"This paper fuses multiple partially blurry color photos of the same scene into one sharp image by treating each pixel's red, green, and blue values as a single quaternion number. The method finds the sharpest regions across the inputs and stitches them together, reporting better scores than eleven earlier fusion methods on three public benchmark sets, though the margins are small.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Focus decisions rely on the unverified monotonicity of lB and lD in Eqs. (14)-(17); without per-patch validation that these measures rank defocus blur correctly across low-gradient, textured, and boundary patches, the claimed state-of-the-art performance of QMCIF is not established.","rationale":"The paper is internally coherent and the authors are candid about metric losses (e.g., QMI on lytro, QY/QCB on mffw and MFI-WHU), which supports a conditional rather than dismissive reading. The strongest claim, however, is the empirical superiority of QMCIF, and that claim rests on the focus maps generated by Eqs. (14)-(17). The measures combine quaternion gradient energies and sparse coefficient energy; the paper provides no derivation showing these are monotone under defocus blur, and the QFED decomposition is nonlinear, so the usual total-variation contractivity argument for convolution does not carry over. Aggregate metrics and three visual crops cannot prove that the measures rank correctly on every patch class; failures in low-gradient or boundary patches would be inherited by the base/detail fusion and by QSSR, whose weights depend on the same lD values. The undisclosed dictionary A and absence of code further compound the issue, but the more load-bearing gap is the missing validation of the focus measure itself. The proposed synthetic-blur per-patch test would settle whether monotonicity holds: if lB/lD rank correctly across all patch classes, the concern is resolved; if not, the SOTA claim requires qualification. Because the paper otherwise provides a full algorithm, convergence analysis, and ablations, the appropriate verdict remains CONDITIONAL as the reader concluded; no change is needed.","tokens_in":18428,"tokens_out":7509,"duration_ms":85600,"concrete_test":"Construct a synthetic ground-truth experiment: take a set of natural all-in-focus color images not in lytro/mffw/MFI-WHU; for each, create focused/blurred pairs by convolving complementary spatial regions, and also entire images, with Gaussian PSFs of several sigma values. Run the QFED module (Eq. 3) as specified, compute lB and lD per patch using Eqs. (14)-(15), and compare the resulting focus maps (Eqs. (16)-(17)) to the known binary ground truth. Report per-class accuracy separately for low-gradient, high-texture, and boundary patches, and plot lB and lD versus sigma for each class. If any class shows non-monotonicity or accuracy well below the aggregate level, the central SOTA claim must be qualified or the focus measure revised. This test requires the authors' pretrained quaternion dictionary A and patch-grouping parameters; those should be released so the check is executable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the empirical superiority of QMCIF, and every fusion decision flows from the dual-scale focus measures in Eqs. (14)-(15) and the threshold maps in Eqs. (16)-(17). The key unstated premise is that lB and lD decrease monotonically as defocus blur increases for all patch classes. This is not a consequence of the QFED model: Eq. (3) is a nonlinear quaternion sparse/low-rank decomposition, so the standard total-variation contractivity of convolution does not transfer to the extracted detail layer D or to the coefficient energy ||z||_2. In low-gradient regions, blur can create spurious small gradients; in textured regions, dictionary coefficients may be redistributed rather than uniformly reduced. No derivation or per-patch experiment is given: Section IV-C and Fig. 6 rely on full-image aggregate metrics (Table III) and selected visual crops, which can hide systematic failures in a minority of patch classes. Moreover, the QSSR refinement (Eq. 22) cannot reliably correct such errors because its weights tau1 and tau2 in Eq. (21) are themselves defined from lD,1 and lD,2; a mislabeled focus measure biases both the initial fusion and the correction. Thus, without evidence that lB and lD rank patches correctly across all color structures, the claimed significant outperformance in Section IV-C is not secured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes QMCIF, a quaternion-domain framework for multi-focus color image fusion. It introduces a quaternion focal element decomposition (QFED) model that separates each color image into a low-rank base layer (via a pretrained quaternion dictionary) and a sparse detail layer, solved by quaternion ADMM with complete update equations. Patch-wise dual-scale focus maps are then generated from base coefficient energy and amplified detail gradient energy, guiding a base-detail fusion strategy; a weighted quaternion structural similarity (WQSSIM) refinement selects final patches. Experiments compare QMCIF with 11 methods on the lytro, mffw, and MFI-WHU datasets using six fusion metrics, and ablation studies examine patch size, fusion stages, adaptive weighting, and quaternion versus real representation. An extension to three-image fusion is demonstrated visually on lytro3 and mffw3.","tokens_in":18672,"tokens_out":5933,"duration_ms":68935,"significance":"If the performance claims are upheld, the paper contributes a self-contained, non-learning quaternion fusion pipeline with explicit optimization steps, broad comparison across three public datasets, and ablations supporting each component. The framework is notable for performing focus detection, fusion, and refinement entirely in the quaternion domain, avoiding per-channel processing. However, the significance is currently moderated by three issues: the central outperformance claim rests on third-decimal metric differences without variance or significance testing; key parameters are selected on the same datasets used for the final comparison; and the focus measures in Eqs. (14)-(17) are assumed to rank defocus blur correctly for all patch classes without per-patch validation. These issues are fixable and do not invalidate the algorithmic contribution, but they must be addressed before the empirical claims can be accepted.","major_comments":[{"comment":"Parameter selection is performed on the same datasets as the final evaluation, which introduces circularity. In Section IV-B, alpha and beta are selected by maximizing QE+QCB on mffw samples, are set to 1.5 and 0.5 for lytro and MFI-WHU based on those datasets, and lambda is chosen using QMI on lytro and mffw. Table III then reports the performance of this tuned configuration on exactly those datasets. Because the 11 competing methods are not given the same per-dataset hyperparameter optimization, part of the reported advantage may be fitted to the evaluation target. The authors should either fix parameters using a separate validation set, report sensitivity across a range of settings, or use nested evaluation, and state clearly which protocol is used.","section":"Section IV-B and Table III"},{"comment":"The claims that QMCIF 'outperforms state-of-the-art methods' and 'significantly outperforms QHOSVD' are not supported by statistical evidence. No standard deviations, per-sample breakdowns, or paired significance tests are reported, and many winning margins are in the third decimal place (e.g., lytro QE: 0.8814 vs. 0.8802 for DBMFIF; mffw QG: 0.7348 vs. 0.7320 for MFIF-GAN). On some metrics QMCIF is not the best (lytro QMI: 1.1656 vs. 1.1920 for MCCSR-Net; MFI-WHU QY: 0.9887 vs. 0.9890 for MGIMF and MFIF-GAN). The authors should provide per-image results and paired tests (e.g., Wilcoxon signed-rank) for the six metrics, and should reserve 'significantly' for differences that are statistically supported.","section":"Section IV-C, Table III"},{"comment":"The focus detection pipeline assumes that the scalar measures lB and lD decrease monotonically as defocus blur increases for all patch classes, including low-gradient, highly textured, and boundary patches. This monotonicity does not follow from the QFED model in Eq. (3), since blur can create spurious gradients in smooth regions and can redistribute dictionary coefficients rather than uniformly reduce them. Because the QSSR weights tau1 and tau2 in Eq. (21) are themselves derived from lD, a mislabeled patch biases both the initial fusion and the refinement correction. The authors should add per-patch validation, for example on synthetically blurred images with known blur radius, stratified by patch type, or provide a formal argument establishing the required monotonicity.","section":"Section III-B2, Eqs. (14)-(17)"},{"comment":"As printed, QSSIM in Eq. (2) is defined as (a)^b with quaternion-valued quantities, which makes QSSIM and hence WQSSIM in Eq. (20) quaternion-valued. Equation (22) then compares WQSSIM values using the ordering operator '>', which is undefined for quaternions. The authors should specify explicitly how QSSIM is reduced to a real scalar (e.g., the standard real-valued formulation of Kolaman and Yadid-Pecht) and how WQSSIM is ranked. Without this clarification, the patch selection rule in Eq. (22) is not well defined.","section":"Definition 5, Eqs. (2), (20), (22)"}],"minor_comments":[{"comment":"In the list of competing methods, 'convolutional sparse coding-bsed' contains a typo and should read 'based'.","section":"Section IV-A"},{"comment":"The text refers to 'GFDF' in the MFI-WHU discussion ('slightly worse than GFDF and MFIF-GAN'), but GFDF is not defined or listed among the 11 comparison methods; please clarify which method this is.","section":"Section IV-C"},{"comment":"The sentence introducing Eq. (20) says 'let f denote a local patch of a base-scale or detail-scale fusion result F2 or F2'; the second 'F2' should presumably be 'F1'.","section":"Section III-D"},{"comment":"The ablation in Table II does not state which dataset it is computed on; the surrounding text should specify the dataset (apparently mffw) and the number of image pairs used.","section":"Section IV-B, Table II"},{"comment":"The conclusion says the framework 'fuses two-scale quaternion matrices and compares them with the input quaternion representations'; this is vague and should be rephrased to describe the actual patch-level WQSSIM selection.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The core algorithmic contribution appears sound: the ADMM updates are fully specified, convergence is shown empirically, and the ablations support the individual components. The main obstacles are empirical rigor (no significance testing, per-dataset parameter tuning) and a few formal gaps in the definitions of QSSIM and WQSSIM. These are addressable in revision, so I do not see a need for rejection. I would also encourage the editor to ask for the code or detailed reproducibility notes, since no implementation is currently provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"QMCIF is a well-built quaternion-domain pipeline for multi-focus color image fusion. What's new is the QFED decomposition (low-rank base via dictionary coding, sparse detail, noise) and the dual-scale focus maps that feed patch selection; nobody has run the whole fusion pipeline in the quaternion domain before. The optimization is fully specified, convergence is shown empirically, and the ablation study supports each component. Credit where due: the paper compares against 11 baselines across three datasets and openly says where it loses (QMI on lytro, QY/QCB on mffw), which is more honest than the field's average.\n\nThe soft spots are real but not fatal. The headline gains are mostly 0.001 to 0.003 on metrics with no variance or significance tests; alpha and beta are chosen on the same datasets where the final scores are reported, so part of the edge is fitted. No code or data is released, and the pretrained quaternion dictionary is not described. The quaternion calculus in Section II has some product-rule and chain-rule identities that look wrong as printed, making the optimization steps hard to verify.\n\nThe stress-test concern about monotonicity of lB and lD under defocus blur is legitimate: the paper assumes these scalar measures rank blur correctly across all patch types, and only aggregate metrics and selected crops are shown. But I'd stop short of calling this load-bearing. Most fusion methods rely on an implicit version of the same assumption; the QFED model at least attempts to separate detail from structure, and the QSSR stage can correct some mislabels. What's missing is a per-patch validation, not a fundamental failure.\n\nNet assessment: plausible, well-engineered, modest novelty. It deserves a serious referee, but as submitted it's not proven state-of-the-art. I'd send it out and require significance testing, out-of-sample parameter selection, and details of the dictionary before accepting. If the authors add those, this is a decent TNNLS paper. If you work on fusion, read it; otherwise a skim suffices.","headline":"A genuine quaternion-domain fusion pipeline with honest ablations, but the SOTA claim rests on tiny margins, per-dataset parameter fitting, and no significance tests.","tokens_in":19323,"tokens_out":2604,"would_cite":false,"duration_ms":30414,"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 a fully quaternion-domain pipeline—quaternion sparse decomposition, dual-scale focus maps, and quaternion structural similarity refinement—produces state-of-the-art multi-focus color image fusion and outperforms the…","keywords":["multi-focus image fusion","color image fusion","quaternion representation","sparse decomposition","focus detection","structural similarity","quaternion domain"],"falsifier":"Synthesize defocused versions of sharp color images by convolving them with known defocus kernels at increasing radii, run QFED on each version, and compare the per-patch values of Eqs. (14) and (15); any patch class in which a more-blurred patch scores higher than the sharp patch would break the focus-map decisions and the whole fusion inherits the error.","tokens_in":18110,"feed_emoji":"🖼️","tokens_out":10803,"duration_ms":118568,"temperature":0.7,"pith_summary":"Multi-focus color image fusion is the task of combining partially focused photos of one scene into a single all-in-focus color image. The paper claims that doing this entirely in quaternion space—where each RGB pixel is one pure quaternion rather than three separate channels—preserves color correlations that grayscale-conversion and channel-wise methods lose. Its framework, QMCIF, learns a quaternion sparse decomposition of each input, builds separate focus maps for base structure and for fine details, fuses the two scales, and then corrects the result patch-by-patch with a quaternion structural similarity measure. The reported experiments on three public benchmarks place QMCIF at or near the top of eleven comparison methods on most fusion metrics, especially edge-structure, phase-consistency, and human-perception scores. If the claim holds, the practical message is that color fidelity in multi-focus fusion can come from the representation itself rather than from post-processing.","feed_headline":"One quaternion pipeline beats 11 rivals at multi-focus color fusion","feed_subtitle":"Keeps color channels coupled through sparse decomposition, dual-scale focus maps, and quaternion structural similarity.","key_machinery":"The load-bearing object is the quaternion focal element decomposition (QFED) of Eq. (3): $\\dot{I} = \\dot{B} + \\dot{D} + \\dot{E}$, where $\\dot{B}$ is a low-rank structural layer over patch groups, $\\dot{D}$ is sparse fine detail, and $\\dot{E}$ is a Gaussian residual, solved iteratively by quaternion ADMM. This model supplies the two quantities that drive everything downstream: the low-rank coefficient matrix $\\dot{Z}$ and the detail layer $\\dot{D}$. Eqs. (14) and (15) turn them into dual-scale focus measures, $l_B = \\|\\nabla_1 \\dot{d}\\|_1 + \\|\\nabla_2 \\dot{d}\\|_1 + \\theta\\|\\dot{z}\\|_2$ and $l_D = \\phi(\\|\\nabla_1 \\dot{d}_s\\|_1 + \\|\\nabla_2 \\dot{d}_s\\|_1)$, whose patch-wise comparisons (Eqs. (16)-(17)) generate the base and detail focus maps. A second mechanism, the weighted quaternion structural similarity (WQSSIM) of Eq. (20), re-scores each candidate fused patch against both sources and selects the better one, correcting regions where the focus maps mislabeled.","core_discovery":"QMCIF's central claim is that multi-focus color fusion should be carried out entirely in the quaternion domain, where the three color channels stay coupled as the imaginary components of one number. The framework decomposes each input quaternion image $\\dot{I}$ into a low-rank structural base layer $\\dot{B}$, a sparse detail layer $\\dot{D}$, and a residual $\\dot{E}$ by solving the quaternion focal element decomposition (QFED) of Eq. (3); from the resulting coefficient matrices and detail layers it builds patch-wise base-scale and detail-scale focus maps. Fusing the two scales separately and then selecting, patch by patch, the candidate whose weighted quaternion structural similarity (WQSSIM) to the sharper source is higher yields the final result. The experiments report that this pipeline achieves the highest fusion performance on the lytro, mffw, and MFI-WHU datasets across most of the six objective metrics, with the clearest gains on phase-congruency ($Q_P$), edge-structural-similarity ($Q_E$), and human-perception ($Q_{CB}$) scores, and that it surpasses the quaternion higher-order-singular-value-decomposition baseline QHOSVD.","pith_inferences":["Not claimed by the paper: the monotonicity of Eqs. (14) and (15) under blur could be tested directly on synthetic defocus and, if it holds, the same measures could serve as a no-reference per-patch focus quality estimator for single images.","Not claimed by the paper: because QFED is a generic color-structure decomposer, its base/detail split could be carried into other color-imaging tasks such as denoising, deblurring, or colorization, where quaternion channel coupling may give the same cross-channel consistency.","Not claimed by the paper: the hand-set parameters (patch size, $\\alpha$, $\\beta$, $\\gamma$) invite a learned or adaptive variant, which could close the remaining gaps to pixel-wise methods on the $Q_{MI}$ and $Q_Y$ metrics."],"forward_implications":["Fusing color images in the quaternion domain, rather than per channel or after grayscale conversion, improves edge-structure and human-perception fusion scores across the lytro, mffw, and MFI-WHU datasets.","Dual-scale focus maps let the method handle both high-texture and low-gradient regions, the failure modes of the singular-value-energy selection used by the quaternion baseline QHOSVD.","The QSSR refinement step corrects mislabeled patches from the initial base/detail fusion, which is what removes ghosting and boundary blur in the reported mffw results.","The framework extends from two-image fusion to fusing three images on the lytro3 and mffw3 sets without changing the core mechanism, producing clean outputs where comparison methods show artifacts.","Ablation experiments indicate that removing any QFED term, fusing at a single scale, replacing adaptive QSSR weights with fixed weights, or dropping the quaternion representation lowers quantitative fusion scores."],"supporting_citations":[{"why":"Prior quaternion-domain MCIF baseline using quaternion higher-order SVD; supplies the singular-value-energy patch selection that QMCIF replaces and the main quaternion rival it reports to outperform.","marker":"[21]"},{"why":"Channel-wise sparse-coding MCIF method; supplies the channel-independent focus measurement and fusion baseline that motivates quaternion coupling.","marker":"[14]"},{"why":"Provides the quaternion soft-thresholding operator used to solve the sparse detail-layer and gradient subproblems.","marker":"[23]"},{"why":"Supplies the quaternion nuclear norm closed-form solution that solves the low-rank subproblem in QFED.","marker":"[25]"},{"why":"Defines quaternion structural similarity (QSSIM), the index that the QSSR refinement strategy weights and maximizes.","marker":"[27]"},{"why":"Provides the mffw dataset of defocus-spread, misaligned multi-focus images used for the main complex-scenario comparisons and ablation.","marker":"[29]"},{"why":"Convolutional sparse representation unfolding network; one of the deep-learning baselines compared in Table III and in visual evaluation.","marker":"[34]"},{"why":"Provides the lytro dataset of well-registered two- and multi-image fusion pairs used for ideal-condition evaluation.","marker":"[35]"},{"why":"Provides the MFI-WHU dataset with small blurred objects used to test small-area focus detection.","marker":"[36]"}],"fun_headline_variants":["Quaternion coupling keeps color channels fused for sharp focus","Sparse quaternion decomposition sharpens multi-focus color fusion","Quaternion base-detail fusion wins on color and detail","All-quaternion fusion tops 11 rivals on multi-focus color"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the two scalar focus measures of Eqs. (14) and (15) decrease monotonically as defocus blur increases for every patch class—low-gradient, highly textured, and boundary regions—so that the sharper patch always wins; the paper does not validate this per patch, only visually and through aggregate metrics.","fun_headline_variants_meta":{"raw":{"variants":["Quaternion coupling keeps color channels fused for sharp focus","Sparse quaternion decomposition sharpens multi-focus color fusion","Quaternion base-detail fusion wins on color and detail","All-quaternion fusion tops 11 rivals on multi-focus color"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000787,"raw_usage":{"total_tokens":3478,"prompt_tokens":958,"completion_tokens":2520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":2451}},"tokens_in":574,"tokens_out":2520,"duration_ms":24562,"temperature":1.0,"reasoning_tokens":2451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:54:28.366121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Synthesize defocused versions of sharp color images by convolving them with known defocus kernels at increasing radii, run QFED on each version, and compare the per-patch values of Eqs. (14) and (15); any patch class in which a more-blurred patch scores higher than the sharp patch would break the focus-map decisions and the whole fusion inherits the error.","supporting_citations":[{"cited_title":"Quaternion higher- order singular value decomposition and its applications in color image processing,","cited_arxiv_id":null,"evidence_quote":"Prior quaternion-domain MCIF baseline using quaternion higher-order SVD; supplies the singular-value-energy patch selection that QMCIF replaces and the main quaternion rival it reports to outperform."},{"cited_title":"Multifocus image fusion and restoration with sparse representation,","cited_arxiv_id":null,"evidence_quote":"Channel-wise sparse-coding MCIF method; supplies the channel-independent focus measurement and fusion baseline that motivates quaternion coupling."},{"cited_title":"Two-dimensional quaternion pca and sparse pca,","cited_arxiv_id":null,"evidence_quote":"Provides the quaternion soft-thresholding operator used to solve the sparse detail-layer and gradient subproblems."},{"cited_title":"Low-rank quaternion approximation for color image processing,","cited_arxiv_id":null,"evidence_quote":"Supplies the quaternion nuclear norm closed-form solution that solves the low-rank subproblem in QFED."},{"cited_title":"Quaternion structural similarity: a new quality index for color images,","cited_arxiv_id":null,"evidence_quote":"Defines quaternion structural similarity (QSSIM), the index that the QSSR refinement strategy weights and maximizes."},{"cited_title":"Unfolding coupled convolutional sparse representation for multi-focus image fusion,","cited_arxiv_id":null,"evidence_quote":"Convolutional sparse representation unfolding network; one of the deep-learning baselines compared in Table III and in visual evaluation."}],"review_version":1}