{"id":"cc123268-a7fc-4144-9758-ff396801bb0c","arxiv_id":"2412.18160","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AMqF is a learned full-reference image quality metric that adaptively decomposes deep features into luminance, contrast, and structure components, encodes them via a learned dictionary of visual words, and scores quality by cosine similarity of dictionary responses.","lead":"The paper proposes an image quality metric, AMqF, that splits deep image features into separate quality factors (brightness, structure, contrast), maps them through a learned dictionary of visual words, and compares reference and distorted images by cosine similarity of the resulting response vectors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global pooling in Eq. 6 discards spatial layout, so AMqF cannot in principle capture regional heterogeneity; the central non-uniform-distortion claim is unsupported.","rationale":"The central claim is that the adaptive quality-factor decomposition and dictionary response improve FR-IQA, especially for non-uniform distortions. My stress-test focuses on the architectural path from response maps to the final score. Equation 6 averages each response map over all spatial positions, collapsing per-region information into a single scalar per dictionary element; Equation 7 then computes cosine similarity between two such global vectors. The method thus cannot distinguish where a distortion occurs, so the repeated claims about 'regional heterogeneity' and 'non-uniformly distorted regions' are inconsistent with the implemented computation. This is an internal inconsistency, not merely a missing baseline. The ablation in Table 3 is also misreported: the text claims removing AMQF reveals the RFDS branch's importance and removing RFDS reveals AMQF's importance, but each removal removes that component, so the interpretation is reversed. Moreover, whole-block ablations cannot attribute gains to spatial pooling versus dictionary learning. A synthetic localized-distortion experiment is the most direct way to settle whether the global histogram can handle region-specific distortions. The reader's weakest assumption about the dictionary is related, but my concern is more fundamental: even with a learned dictionary, global pooling prevents regional modeling. The CONDITIONAL verdict remains appropriate, but the revision must add the spatial-variant and localized-distortion test; hence the verdict is unchanged.","tokens_in":12191,"tokens_out":9704,"duration_ms":88165,"concrete_test":"Construct a synthetic localized-distortion experiment: for 100 pristine images, generate paired distorted images with matched global distortion energy (e.g., Gaussian blur applied either uniformly or only to a quadrant). Run (i) the full AMqF, (ii) a variant that replaces global average pooling with spatial pyramid pooling of the response maps, and (iii) a variant with a fixed random dictionary. If AMqF does not score the localized distortion differently from the uniform one, or if the spatial variant outperforms it, the regional-heterogeneity claim is falsified. If the random-dictionary variant matches AMqF, the learned dictionary is not the source of the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims AMqF captures non-uniform distortions via regional heterogeneity (Abstract, Section 4.1). However, the final quality score (Eq. 7) is the cosine similarity between global histograms Pref and Pdist obtained by averaging response maps over all spatial positions (Eq. 6). This discards spatial layout: a distortion confined to a small region and a uniform distortion producing the same visual-word frequency histogram yield indistinguishable scores. The architecture therefore cannot represent 'regional heterogeneity' or 'non-uniformly distorted regions' as claimed, so any observed improvement over SOTA cannot be attributed to the stated regional mechanism. The ablation in Table 3 does not resolve this because it removes entire blocks (AMQF/RFDS) rather than isolating spatial pooling, and the accompanying text misattributes the direction of the effect (removing AMQF is said to implicate the RFDS branch, and vice versa). No experiment distorts only a subset of image regions, leaving the central non-uniform-distortion claim untested.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AMqF, a full-reference IQA method that decomposes deep features from a pretrained backbone into multiple quality factors (luminance, contrast, structure) via an adapter, quantizes these factors into discrete visual words in a learnable dictionary space, and measures visual similarity via cosine similarity between global average response histograms of reference and distorted images. Experiments on LIVE, CSIQ, TID2013, and KADID-10k report competitive PLCC/SRCC values, with the best results on TID2013 and KADID-10k and top-three performance on LIVE and CSIQ. The abstract and introduction claim that the dictionary-based response captures regional heterogeneity and non-uniform distortions, but this central claim is not supported by the experimental design or the scoring mechanism.","tokens_in":12408,"tokens_out":7388,"duration_ms":60928,"significance":"If validated, the method would offer a novel angle for FR-IQA by introducing an adaptive factor decomposition and a dictionary-response representation. Strengths include a broad comparison with 19 methods across four datasets, a cross-database evaluation, and an anonymous code release, all of which are positive features. However, the claim of outperforming state-of-the-art is only partially supported by Table 1, and the proposed mechanism for non-uniform distortions is contradicted by the global pooling in Eq. (6) and is never tested directly. The ablation analysis in Section 5.3 contains misattributions that further weaken the evidence. The contribution is potentially interesting but requires substantial additional validation.","major_comments":[{"comment":"The final quality score in Eq. (7) is computed as the cosine similarity between global average response vectors Pref and Pdist, where each response map R_k is averaged over spatial positions in Eq. (6). This averaging discards all spatial layout information. As a result, a distortion confined to a small region and a uniform distortion that yield the same average response histogram produce identical scores, so the model cannot in principle distinguish them. The abstract and Section 4.1 claim that AMqF precisely captures quality features in non-uniformly distorted regions, but the scoring function does not use any spatial information. To support the central claim, the scoring must either preserve spatial layout or the paper must provide explicit experiments with localized distortions showing that the method behaves differently from global-histogram matching.","section":"Section 4.3, Eqs. (5)-(7)"},{"comment":"The statement that AMqF 'outperforms existing state-of-the-art approaches' is not fully supported by Table 1. On LIVE, TOPIQ-FR achieves PLCC/SRCC of 0.984/0.984 compared with 0.979/0.980 for AMqF; on CSIQ, TOPIQ-FR achieves 0.980/0.978 compared with 0.975/0.974. AMqF is best only on TID2013 and KADID-10k, and TOPIQ-FR is not evaluated on KADID-10k. The claim should be qualified as 'competitive with or better than SOTA on specific datasets.' In addition, the AMqF-VGG rows in Table 1 are never mentioned in the text, and the color-based highlighting of best/second/third results is not decipherable in the printed version; the table should explicitly mark ranking.","section":"Section 5.2.1 and Table 1"},{"comment":"No experimental protocol details are reported. The paper does not state the training/validation split on reference images (e.g., the standard 80/20 split), the number of random runs, or the standard deviations of PLCC/SRCC. Without these, the reported improvements over TOPIQ-FR on TID2013 (0.968 vs 0.958 PLCC) and over JND-SalCAR on KADID-10k (0.964 vs 0.960) cannot be assessed for statistical significance. The cross-database experiment in Table 2 trains on the entire KADID-10k database, which is a different protocol from the main experiments; the relationship between the two protocols should be clarified.","section":"Section 5.1 and Tables 1-3"},{"comment":"The ablation discussion misattributes the direction of the effects. The text states that removing AMQF 'indicates that the RFDS branch plays a crucial role,' but removing AMQF leaves the RFDS branch alone; similarly, removing RFDS is stated to demonstrate that AMQF enhances performance, whereas it leaves AMQF alone. The numbers also undermine the claim of complementarity: AMQF alone achieves LIVE PLCC 0.978, CSIQ PLCC 0.971, TID2013 PLCC 0.968, and KADID PLCC 0.947, close to the full model, whereas RFDS alone drops to 0.936, 0.970, 0.871, and 0.947 respectively. The marginal contribution of RFDS is therefore small or negative on some datasets, and the claimed 'indispensability' is not supported.","section":"Section 5.3 and Table 3"},{"comment":"There are index errors that prevent reproduction. In Eq. (5), v_k is indexed by spatial positions (m,n), but V is defined as an N×D matrix in Eq. (4) with D the feature dimension; if v_k is a D-dimensional vector, the summation over C, H, W with v_k[m,n] is mismatched. In Eq. (6), the average for Pdist[k] and Pref[k] is taken only over i, leaving j free, and the symbol N is used for both the number of dictionary words (Eq. 4) and the number of spatial positions. The intended global average pooling should sum over both spatial indices and should use distinct notation for the two quantities.","section":"Section 4.3, Eqs. (5)-(6)"},{"comment":"The core mechanism of a learned dictionary space is not validated. There is no comparison against a fixed or randomly initialized dictionary, no analysis of which visual words respond to which distortions, and no ablation that isolates the dictionary learning from the rest of the architecture. Without these controls, the paper does not demonstrate that the 'comprehensive and discriminative dictionary space' contributes to the reported performance.","section":"Section 4.3"}],"minor_comments":[{"comment":"The notation Iref and Idist appears swapped: the text introduces 'given a batch of distorted images Iref and reference images Idist', but then uses Iref as the reference and Idist as the distorted input in the following equations. Please make the notation consistent throughout.","section":"Section 4.2"},{"comment":"The decoder for single-channel image reconstruction is described only qualitatively; no architecture, no training schedule, and no weight for the reconstruction loss Lre in the total loss are given. This makes the contribution of the reconstruction constraint unverifiable.","section":"Section 4.2"},{"comment":"The labels Lref, Cref, Sref and Ldist, Cdist, Sdist are not defined in the caption, and the flow from the adapter to the dictionary and the score is hard to follow; a clearer diagram would help.","section":"Figure 1"},{"comment":"References [5] and [6] are duplicated (same paper by Cheon et al.), and reference [7] lists the same author twice; these should be corrected.","section":"References"},{"comment":"The connection to the HVS is described qualitatively and is not operationalized in the model; for example, no parameter or constraint in the adapter is derived from HVS principles, so the claim that the adaptive factors 'align with human visual perception' is not empirically evidenced beyond the overall IQA accuracy.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a competitive FR-IQA method with a novel dictionary-response framework, but the central non-uniform-distortion claim is not supported by the architecture (global pooling) and is not tested. The missing experimental details (split, error bars) and the ablation misattribution are fixable, but the spatial-layout issue may require a redesign of the scoring or at least dedicated localized-distortion experiments. I recommend major revision rather than rejection because the core idea is worth exploring and the deficiencies are addressable. The comparison with TOPIQ-FR should also be presented more honestly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The strongest thing in this paper isn't the headline claim about regional heterogeneity—it's the cross-database result in Table 2. Training on KADID-10k and testing on LIVE/CSIQ/TID2013 gives a real jump over WaDIQaM-FR and RADN (0.907 vs 0.837 PLCC on LIVE). That's a concrete, useful result, and the fact that they release code is a point in their favor.\n\nThe architecture is a plausible integration of familiar pieces: ResNet50 features, an adapter that decomposes into luminance/contrast/structure factors, a reconstruction loss, and a learned dictionary whose cosine-similarity score produces the quality prediction. It's not a new paradigm, but the combination is coherent and the KADID-10k/TID2013 numbers are competitive.\n\nNow the soft spots, in order of severity. First, the central mechanism is unvalidated and, as written, cannot do what the authors claim. The final score in Eq. 7 is a cosine similarity between global average-pooled response vectors (Eq. 6). Global averaging discards all spatial layout: an image with a small patch of distortion and an image with the same visual-word frequencies spread uniformly across the whole frame will produce the same score. The paper repeatedly claims to capture \"non-uniform distortions\" and \"regional heterogeneity,\" but nothing in the method actually encodes spatial arrangement. The ablation doesn't help—removing entire blocks doesn't isolate spatial pooling, and the text misattributes the direction of the effects (removing AMQF is said to implicate the RFDS branch, and vice versa).\n\nSecond, the SOTA claim is overstated. In Table 1, TOPIQ-FR is clearly better on LIVE and CSIQ; AMqF wins only on TID2013 and KADID-10k. The abstract and Section 5.2.1 say it \"outperforms existing state-of-the-art,\" which is not supported by their own table.\n\nThird, there are small but telling technical errors: Eqs. 5 and 6 have index/notation inconsistencies (the convolution sum over m,n with a 1D dictionary vector, and the average pooling missing the j-index). No error bars, no train/validation split details, no statistical comparison of the small differences.\n\nNone of this destroys the paper's potential. The cross-database generalization is genuinely interesting, and the method may work for reasons unrelated to the stated regional mechanism. But the current version would mislead a reader who trusts the abstract.\n\nRecommendation: send it to peer review. A good referee should push the authors to (1) either add a spatial-pooling variant that actually preserves layout or stop claiming to capture regional heterogeneity, (2) correct the SOTA claim, (3) fix the equations, and (4) provide error bars and split details. With those changes, it could be a solid contribution.","headline":"A competitive FR-IQA baseline with real cross-database results, but the regional-heterogeneity story is not supported by the architecture's global pooling.","tokens_in":12938,"tokens_out":3214,"would_cite":false,"duration_ms":30229,"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":"A full-reference image quality metric that decomposes deep features into luminance, contrast, and structure and scores their responses in a learned dictionary space claims state-of-the-art agreement with human opinion on TID2013 and…","keywords":["full-reference image quality assessment","adaptive quality factors","dictionary space","visual words","regional heterogeneity","non-uniform distortion","deep feature similarity","image quality metric"],"falsifier":"Train the same pipeline but replace the learned dictionary with a fixed random basis of the same size, or freeze the dictionary at initialization, and compare PLCC/SRCC on TID2013 and KADID-10k; if the scores remain near $0.968/0.968$ and $0.964/0.961$, the claimed dictionary mechanism is not what drives the result.","tokens_in":12012,"feed_emoji":"🖼️","tokens_out":6412,"duration_ms":56279,"temperature":0.7,"pith_summary":"The paper proposes a full-reference image quality metric that measures perceptual similarity by decomposing deep features into luminance, contrast, and structure components, then projecting those components onto a learnable dictionary of 1,024 visual words. The response of each quality factor to the dictionary basis vectors becomes a coordinate vector, and the quality score is the cosine similarity of reference and distorted coordinate distributions. On TID2013 and KADID-10k the reported correlations with human opinion scores are the best among the compared methods, and on LIVE and CSIQ the method stays in the top three. The work is aimed at distortions that vary across regions, where global metrics such as SSIM assume uniformity.","feed_headline":"Image-quality metric beats prior methods on two hard benchmarks","feed_subtitle":"Decomposing images into luminance, contrast, and structure, then scoring their dictionary responses, beats previous metrics on two datasets.","key_machinery":"The load-bearing object is the learned dictionary basis matrix $V \\in \\mathbb{R}^{1024 \\times 512}$ together with the convolution-based response operation that projects each normalized quality-factor feature map into a distribution over 1,024 visual words. The paper calls this the response of factors in dictionary space: for each visual word $k$, a response map $R_k[i,j]$ is computed, then average pooling condenses it to a scalar per image, and cosine similarity between the reference and distorted response distributions gives the quality score. An adapter decomposes deep features into luminance, contrast, and structure factors, and reconstruction and decorrelation losses shape those factors during training.","core_discovery":"The central claim is that representing image quality as the response of adaptively selected quality factors in a learned dictionary space captures regional heterogeneity better than global deep-feature distances. With the AMqF pipeline, deep features are split into multi-head and luminance/contrast/structure streams, strengthened by a single-channel reconstruction decoder, and projected onto dictionary basis vectors $V\\in\\mathbb{R}^{1024\\times512}$ by convolution; average-pooled responses form distributions $P_{\\text{ref}}$ and $P_{\\text{dist}}$ whose cosine similarity is the predicted quality. The paper reports PLCC/SRCC of $0.968/0.968$ on TID2013 and $0.964/0.961$ on KADID-10k, exceeding the compared methods, and top-three results on LIVE and CSIQ, arguing this reflects better handling of non-uniformly distorted regions.","pith_inferences":["Implicit testable extension: a per-distortion correlation breakdown on TID2013 and KADID-10k would show whether the advantage is actually concentrated in non-uniform and mixed distortions, as the paper argues.","Editorial reading: the dictionary response is conceptually a bag-of-visual-words histogram over deep-feature patches; if the gains reproduce, the same quantized-response idea could transfer to other perceptual similarity and retrieval tasks.","Editorial reading: because the adapter, reconstruction decoder, and dictionary are introduced together, the aggregate numbers do not isolate the dictionary's contribution; a variant that keeps everything fixed and swaps only the learned dictionary for a fixed random one would test the claimed mechanism directly."],"forward_implications":["On TID2013 the method reports PLCC/SRCC of $0.968/0.968$, the best among the compared methods on that benchmark.","On KADID-10k it reports PLCC/SRCC of $0.964/0.961$, also the best among the compared methods.","On LIVE and CSIQ the reported correlations stay within the top three, so the advantage is not confined to one database.","Trained on KADID-10k and tested on other databases without fine-tuning, the model improves over the compared cross-database baselines on LIVE, CSIQ, and TID2013.","Removing either the adaptive quality-factor branch or the dictionary-response branch lowers performance, so both components are claimed to contribute to the final scores."],"supporting_citations":[{"why":"Defines the structural-similarity baseline and the luminance/contrast/structure decomposition that AMqF adapts.","marker":"[39]"},{"why":"Supplies the deep-feature perceptual-distance baseline LPIPS that AMqF compares against and aims to improve.","marker":"[44]"},{"why":"TOPIQ-FR is the strongest compared deep FR-IQA method on TID2013 that AMqF must beat.","marker":"[4]"},{"why":"DISTS provides a deep structure-and-texture similarity baseline for comparison.","marker":"[9]"},{"why":"A-DISTS is the locally adaptive deep structure/texture baseline for comparison.","marker":"[10]"},{"why":"KADID-10k supplies the training set and one of the two test benchmarks where AMqF reports the best results.","marker":"[23]"},{"why":"TID2013 supplies the test benchmark where AMqF reports the best correlations.","marker":"[28]"},{"why":"LIVE supplies one of the test benchmarks where AMqF reports top-three results.","marker":"[32]"},{"why":"CSIQ supplies one of the test benchmarks where AMqF reports top-three results.","marker":"[20]"}],"fun_headline_variants":["Adaptive multi-quality factors outperform on TID2013 and KADID","Dictionary-space IQA captures regional distortion better than rivals","New metric uses adaptive quality factors to win on two benchmarks","Regional heterogeneity tackled by adaptive quality factors in dictionary space","Adaptive quality factors in dictionary space beat previous IQA methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a trainable dictionary of 1,024 basis vectors, initialized randomly and learned from distorted training images, creates a complete and discriminative space that generalizes to unseen distortions; the paper does not compare it against a fixed or random dictionary, so the dictionary's specific role is what carries the claimed improvement.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive multi-quality factors outperform on TID2013 and KADID","Dictionary-space IQA captures regional distortion better than rivals","New metric uses adaptive quality factors to win on two benchmarks","Regional heterogeneity tackled by adaptive quality factors in dictionary space","Adaptive quality factors in dictionary space beat previous IQA methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1550,"prompt_tokens":950,"completion_tokens":600,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":516}},"tokens_in":566,"tokens_out":600,"duration_ms":5963,"temperature":1.0,"reasoning_tokens":516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:58:16.376567+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline but replace the learned dictionary with a fixed random basis of the same size, or freeze the dictionary at initialization, and compare PLCC/SRCC on TID2013 and KADID-10k; if the scores remain near $0.968/0.968$ and $0.964/0.961$, the claimed dictionary mechanism is not what drives the result.","supporting_citations":[{"cited_title":"Image quality assessment: from error visibility to structural similarity","cited_arxiv_id":null,"evidence_quote":"Defines the structural-similarity baseline and the luminance/contrast/structure decomposition that AMqF adapts."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-feature perceptual-distance baseline LPIPS that AMqF compares against and aims to improve."},{"cited_title":"Topiq: A top-down approach from semantics to distortions for image quality assessment","cited_arxiv_id":null,"evidence_quote":"TOPIQ-FR is the strongest compared deep FR-IQA method on TID2013 that AMqF must beat."},{"cited_title":"Image quality assessment: Unifying structure and texture similarity","cited_arxiv_id":null,"evidence_quote":"DISTS provides a deep structure-and-texture similarity baseline for comparison."},{"cited_title":"Locally adaptive structure and texture similarity for image quality assessment","cited_arxiv_id":null,"evidence_quote":"A-DISTS is the locally adaptive deep structure/texture baseline for comparison."},{"cited_title":"Kadid-10k: A large-scale artificially distorted iqa database","cited_arxiv_id":null,"evidence_quote":"KADID-10k supplies the training set and one of the two test benchmarks where AMqF reports the best results."},{"cited_title":"Color image database tid2013: Peculiarities and preliminary re- sults","cited_arxiv_id":null,"evidence_quote":"TID2013 supplies the test benchmark where AMqF reports the best correlations."},{"cited_title":"A statistical evaluation of recent full reference image quality assessment algorithms","cited_arxiv_id":null,"evidence_quote":"LIVE supplies one of the test benchmarks where AMqF reports top-three results."},{"cited_title":"Most apparent distortion: full-reference image quality assessment and the role of strategy","cited_arxiv_id":null,"evidence_quote":"CSIQ supplies one of the test benchmarks where AMqF reports top-three results."}],"review_version":1}