{"id":"f2249e94-9eec-428d-945d-c54b4282851c","arxiv_id":"1908.01703","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SESF-Fuse fuses multi-focus images by computing spatial frequency on deep features from an unsupervised autoencoder, then using a refined decision map to blend the source images.","lead":"A team from USTB proposes an unsupervised deep-learning method for multi-focus image fusion: an autoencoder extracts deep feature maps, and a spatial-frequency measure on those maps decides which regions are in focus. The method beats sixteen baselines on standard fusion metrics, though the gains are small and statistical significance is not demonstrated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"State-of-the-art claim rests on tiny mean-metric gaps with no significance testing; the advantage over CNN-Fuse (ΔQg=0.0003, ΔQcb=0.0009) may be noise.","rationale":"The reader correctly identified generalization and parameter tuning as weaknesses, but the most load-bearing issue is that the headline SOTA conclusion is not supported by any statistical evidence. Table 2 shows SESF-Fuse and CNN-Fuse are separated by sub-0.001 differences in Qg and Qcb, while SESF-Fuse is worse on Qm. Without per-image variance or paired tests, these differences could be noise, and the first-place counts are not a rigorous test. This directly undermines the strongest claim, independent of whether the deep-feature spatial-frequency hypothesis is correct or whether post-processing parameters were tuned. I therefore keep the conditional verdict: the paper is plausible and has a public implementation, but the SOTA assertion needs a significance analysis and ideally held-out validation. My agreement with the reader is partial because the reader focused on the deep-feature reliability and parameter tuning, whereas I see the lack of significance testing as the primary obstacle; the two concerns compound rather than replace each other.","tokens_in":8866,"tokens_out":6099,"duration_ms":56723,"concrete_test":"Reproduce Table 2 using the released code on the same 38 test pairs and compute per-image Qg, Qm, and Qcb for SESF-Fuse and CNN-Fuse (and the other top methods). Apply a paired Wilcoxon signed-rank test and a paired bootstrap with 10,000 resamples to the per-image differences, reporting p-values and 95% confidence intervals for ΔQg and ΔQcb. If the confidence interval includes zero or p≥0.05, the state-of-the-art claim should be downgraded to 'competitive with CNN-Fuse.' As a secondary check, hold out half of the pairs, tune the guided-filter parameters (r, ε) and area threshold on the other half, and evaluate on the held-out half; if the ranking changes, the reported parameters were overfit to the test set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SESF-Fuse achieves state-of-the-art multi-focus fusion performance, and Table 2 is the sole quantitative evidence. The decisive margins are extremely small: Qg is 0.7105 versus 0.7102 for CNN-Fuse, Qcb is 0.7848 versus 0.7839, and Qm is actually lower (2.8886 versus 2.9654 for CNN-Fuse and 2.9642 for LP). The paper reports only means and first-place counts over 38 test pairs, with no per-image variance, confidence intervals, or paired significance tests. First-place counts are not a substitute for a significance test and can be dominated by near-ties. For the state-of-the-art claim to hold, the observed Qg and Qcb advantages must be systematic; at margins of roughly 0.0003 to 0.0009, one or two favorable images could flip the result. The reader's concern about post-processing parameters tuned on the test set is related, but the more fundamental problem is that even with fixed parameters, the reported evidence does not establish a statistically reliable advantage. The conditional verdict is therefore appropriate unless a proper significance analysis is supplied.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SESF-Fuse, an unsupervised encoder-decoder model for multi-focus image fusion. The network is trained on MS-COCO grayscale images to reconstruct inputs, and at inference time spatial frequency computed on deep features is used to build an initial decision map, followed by morphological consistency verification and guided-filter refinement. The fused image is obtained by pixelwise weighted averaging of the two source images according to the refined decision map. The authors compare against 16 existing fusion methods on 38 multi-focus image pairs using three objective metrics (Qg, Qm, Qcb) and qualitative visualization, and claim state-of-the-art performance based on the highest mean Qg and Qcb values and the largest number of first-place results.","tokens_in":9132,"tokens_out":2631,"duration_ms":26788,"significance":"If the claimed state-of-the-art result holds, the paper makes a useful contribution by demonstrating that feature-gradient activity measures computed on unsupervised deep features can outperform supervised and hand-crafted fusion rules for multi-focus fusion. Strengths of the work include the unsupervised training protocol on independent MS-COCO data, the clear ablation study separating the contributions of spatial-frequency fusion and decision-map refinement, the availability of code, and the use of standard objective metrics and a large comparison set. However, the statistical and generalization evidence for the central claim is currently insufficient, and several fusion hyperparameters appear to be chosen on the test set used for the headline comparisons.","major_comments":[{"comment":"The state-of-the-art claim rests on mean differences of Qg=0.7105 versus 0.7102 for CNN-Fuse and Qcb=0.7848 versus 0.7839 for CNN-Fuse, i.e., margins of 0.0003 and 0.0009 over 38 test pairs. No per-image variance, confidence intervals, or paired significance tests are reported, so these margins may be within noise. Please provide per-image scores and a paired significance test (e.g., Wilcoxon signed-rank test) for Qg, Qcb, and Qm, and report effect sizes, before claiming superiority over CNN-Fuse.","section":"Experiments, Table 2"},{"comment":"Several fusion hyperparameters are explicitly described as empirically set: the guided filter parameters r=4 and ε=0.1, the small-region area threshold 0.01×H×W, and the structuring-element radius tied to the spatial-frequency kernel radius. If these values were selected using the same 38 test pairs that appear in Table 2, the reported metrics are optimistically biased. Please report how these parameters were chosen, use a separate validation set or cross-validation, and include a sensitivity analysis to demonstrate that the results are not artifacts of this tuning.","section":"Method, Fusion subsection; Experiments"},{"comment":"The proposed method's mean Qm is 2.8886, which is lower than CNN-Fuse (2.9654) and LP (2.9642). The paper argues that the largest first-place count (14 versus 7 and 8) makes the method more robust, but first-place counts do not account for the magnitude of differences or near-ties and are not a substitute for a statistical comparison of the Qm distributions. The concluding claim that the method 'generally outperform[s] other methods, leading to state-of-the-art performance' is therefore not supported on all three metrics.","section":"Experiments, Table 2 and Conclusion"}],"minor_comments":[{"comment":"The phrase 'In order to access the fusion performance' should be 'assess the fusion performance'.","section":"Experiments, Objective Image Fusion Quality Metrics"},{"comment":"The text inconsistently uses 'state-of-art' and 'state-of-the-art'; please standardize to 'state-of-the-art'.","section":"Abstract and throughout"},{"comment":"'The key point behind of proposed method' should be 'behind the proposed method'.","section":"Abstract"},{"comment":"The sentence 'we use 38 pairs of multi-focus images as testing set for evaluation, which are publicly available online' does not name the specific datasets or sources; please list them precisely for reproducibility.","section":"Experiments, Experimental Settings"},{"comment":"The author name 'Savi´c' appears with a malformed accent; please use proper Unicode or LaTeX encoding.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal, but the central claim of state-of-the-art performance currently rests on statistically unverified margins and test-set-tuned hyperparameters. The authors should be asked to supply significance testing and an unbiased parameter-selection protocol, or to temper the claim accordingly. The first-place count metric should be interpreted with caution, as it can be dominated by near-ties."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a modest, honest incremental paper, and the headline claim about state-of-the-art performance doesn't survive contact with the numbers. The genuinely new idea is to compute spatial frequency on deep features from an unsupervised encoder-decoder and use that to build a decision map, instead of fusing features directly like DenseFuse. That is a reasonable extension, and the ablation supports the core hypothesis: the 'se sf dm' configuration clearly beats feature-intensity fusion modes (max, average, L1-norm) on the three reported metrics. Credit where it's due: training on MS-COCO is independent, the code is released, and the comparison against 16 methods on 38 pairs is useful.\n\nThe soft spot is the evidence for the central claim. The margins over CNN-Fuse on Qg and Qcb are 0.0003 and 0.0009, which is noise-level without per-image variance or paired significance testing. First-place counts do not fix that, especially since the method actually loses on mean Qm to LP and CNN-Fuse. The paper acknowledges this and explains it, which is honest, but the abstract still says 'state-of-the-art.' Related concern: the guided-filter parameters are 'experimentally set' on the same 38 test pairs, so the reported numbers are likely optimistic even if they are reproducible. This is not a fatal flaw in the method itself; it is a flaw in the strength of the claim.\n\nThe paper is best read as a useful comparison point for researchers working on multi-focus fusion. It shows that a simple unsupervised activity measure on deep features can compete with a supervised classifier, which is interesting for the field. But the method's reported superiority is not established, and the authors should be asked for held-out validation, error bars, and a significance test before the SOTA claim is taken seriously.\n\nMy take: send it to peer review, but reject the headline claim in the abstract unless the authors can show the tiny gaps are systematic. The underlying contribution deserves referee time; the current evidence does not deserve the conclusion.","headline":"A plausible unsupervised twist on multi-focus fusion whose 'state-of-the-art' claim is not backed by the evidence; worth a serious referee, but not a reliable SOTA result as written.","tokens_in":9639,"tokens_out":1765,"would_cite":true,"duration_ms":21625,"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":"This paper claims that measuring spatial frequency on deep features, rather than feature intensity, gives state-of-the-art multi-focus image fusion from an unsupervised encoder-decoder, with the highest reported mean Qg (0.7105) and Qcb…","keywords":["multi-focus image fusion","unsupervised deep learning","encoder-decoder network","spatial frequency","deep feature gradient","decision map","consistency verification","image fusion metrics"],"falsifier":"Use a multi-focus pair with a known ground-truth focus mask and compare the decision maps from deep-feature spatial frequency versus raw grayscale spatial frequency under identical post-processing; if raw features match or beat the deep version, the deep encoder is not the source of the advantage. Likewise, if randomly reinitializing the encoder leaves the three metrics essentially unchanged, the learned reconstruction features are not what the method relies on.","tokens_in":8691,"feed_emoji":"🖼️","tokens_out":6789,"duration_ms":65393,"temperature":0.7,"pith_summary":"This paper is trying to establish that, for multi-focus image fusion, the right way to use an unsupervised deep network is to measure gradient sharpness (spatial frequency) on the learned features rather than feature intensity. If true, it means a network trained only to reconstruct ordinary images can supply a focus detector, removing the need for synthetic blurred training pairs. On 38 multi-focus test pairs, the method reports the highest average Qg (0.7105) and Qcb (0.7848) among 16 comparison methods, and the greatest number of per-image first places. A sympathetic reader would take the contribution as evidence that deep-feature gradient activity, followed by consistency verification, is a strong and general fusion rule.","feed_headline":"Deep-feature sharpness picks which pixels are in focus","feed_subtitle":"Unsupervised spatial frequency on learned features beats 16 fusion methods on 38 image pairs.","key_machinery":"The carrying object is a pixel-wise spatial frequency computed on deep feature vectors, not on gray levels. For each position the row and column frequencies are summed over a radius-r neighborhood of the feature map and combined as $SF(x,y)=\\sqrt{CF(x,y)^2+RF(x,y)^2}/(2r+1)^2$; the decision map is simply which source has the larger $SF$ at each pixel. This expresses the paper's assumption that objects within the depth of field are sharp and therefore have larger feature gradients. The map is then cleaned by alternating opening and closing with a disk structuring element, small-region reversal below $0.01\\times H\\times W$, and an edge-preserving guided filter, before the weighted-average fusion step.","core_discovery":"The central claim, stated on the paper's own terms, is that activity level in multi-focus fusion should be measured by feature gradient rather than feature intensity. The paper builds an encoder-decoder trained only to reconstruct grayscale MS-COCO images, extracts deep feature maps, computes a pixel-wise spatial frequency on those maps, and thresholds the comparison to form a focus decision map. Consistency verification, through morphological opening and closing, removal of regions below 0.01 by image height and width, and guided filtering, refines the map, and the fused image is a weighted average of the two sources. The reported result is state-of-the-art fusion performance versus 16 methods, with the highest mean Qg and Qcb and the most first-place finishes across 38 image pairs.","pith_inferences":["If the deep-feature spatial-frequency measure generalizes, it could serve as a focus cue for focal-stack depth estimation in microscopy and computational photography, where no paired blurred images exist.","Because the network was trained on generic photo content, the method's performance on non-photographic imagery such as medical scans, text, or synthetic graphics is untested; a cross-domain evaluation would bound its generality.","The comparison is summarized by three metric averages, and since SESF-Fuse is below LP on mean Qm, a per-image paired significance test over the 38 pairs would settle whether the overall state-of-the-art claim holds under a different aggregation.","A controlled sensitivity analysis of the hand-set parameters (disk radius, area threshold, guided filter radius and regularization) is a natural next experiment; stable rankings across a wide parameter range would make the method genuinely parameter-free in practice."],"forward_implications":["Multi-focus fusion can be done without synthetic blurred training pairs: a network trained only to reconstruct ordinary grayscale images supplies the focus cue.","Gradient-based activity on deep features is a better fusion rule than intensity-based L1-norm, max, average, or abs-max rules, according to the paper's ablation study.","The consistency-verification post-processing converts noisy per-pixel focus decisions into clean region boundaries, which is why the decision-map variant outperforms direct deep-feature fusion.","The scheme naturally extends to any number of source images by serial pairwise fusion.","The same deep-feature spatial-frequency recipe is suggested as applicable to multi-exposure, infrared-visible, and medical image fusion."],"supporting_citations":[{"why":"Supplies the encoder-decoder unsupervised architecture and the L1-norm fusion baseline that the paper adapts and compares against.","marker":"[Li and Wu 2019]"},{"why":"Supplies the supervised CNN fusion baseline and the small-region removal consistency strategy reused here.","marker":"[Liu et al. 2017]"},{"why":"Defines the original spatial frequency on gray images, which the paper extends to deep feature maps.","marker":"[Li, Kwok, and Wang 2001]"},{"why":"Provides the morphological opening and closing operations used to smooth the initial decision map.","marker":"[De, Chanda, and Chattopadhyay 2006]"},{"why":"Provides the guided filter used to refine decision-map boundaries in the consistency verification stage.","marker":"[He, Sun, and Tang 2013]"},{"why":"Provides the MS-COCO training images used for the unsupervised reconstruction training of the encoder-decoder.","marker":"[Lin et al. 2014]"},{"why":"Defines the Qg objective fusion metric used as one of the three primary evaluation measures.","marker":"[Xydeas and Petrovic 2000]"},{"why":"Defines the Qcb objective fusion metric used to compare fusion quality.","marker":"[Chen and Blum 2009]"},{"why":"Defines the Qm objective fusion metric used as the third evaluation measure.","marker":"[Peng-wei Wang and Bo Liu 2008]"},{"why":"Supplies the SSIM loss component used to train the encoder-decoder reconstruction network.","marker":"[Wang et al. 2004]"}],"fun_headline_variants":["Unsupervised deep fusion uses spatial frequency on learned features","Deep-feature spatial frequency outdoes 16 methods in focus fusion","SESF-Fuse: unsupervised multi-focus fusion beats 16 existing methods","Unsupervised encoder-decoder uses feature sharpness for focus fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that spatial frequency computed on deep features from a network trained on unrelated grayscale photos reliably measures focus on new multi-focus pairs, and that the hand-set post-processing parameters do not quietly account for the reported gains.","fun_headline_variants_meta":{"raw":{"variants":["Unsupervised deep fusion uses spatial frequency on learned features","Deep-feature spatial frequency outdoes 16 methods in focus fusion","SESF-Fuse: unsupervised multi-focus fusion beats 16 existing methods","Unsupervised encoder-decoder uses feature sharpness for focus fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3281,"prompt_tokens":828,"completion_tokens":2453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":2377}},"tokens_in":444,"tokens_out":2453,"duration_ms":17953,"temperature":1.0,"reasoning_tokens":2377,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:05:09.004954+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a multi-focus pair with a known ground-truth focus mask and compare the decision maps from deep-feature spatial frequency versus raw grayscale spatial frequency under identical post-processing; if raw features match or beat the deep version, the deep encoder is not the source of the advantage. Likewise, if randomly reinitializing the encoder leaves the three metrics essentially unchanged, the learned reconstruction features are not what the method relies on.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the supervised CNN fusion baseline and the small-region removal consistency strategy reused here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MS-COCO training images used for the unsupervised reconstruction training of the encoder-decoder."},{"cited_title":"S., and Petrovic , V","cited_arxiv_id":null,"evidence_quote":"Defines the Qg objective fusion metric used as one of the three primary evaluation measures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Qcb objective fusion metric used to compare fusion quality."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Qm objective fusion metric used as the third evaluation measure."},{"cited_title":"C.; Sheikh, H","cited_arxiv_id":null,"evidence_quote":"Supplies the SSIM loss component used to train the encoder-decoder reconstruction network."}],"review_version":1}