{"id":"fdf37a37-9e3d-4b58-8fb2-eeccf0aba189","arxiv_id":"2502.05995","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A brief categorical review of low-light image enhancement methods that summarizes about ten pipelines but provides no experiments, no comparison tables, and no systematic literature search.","lead":"This paper is a short survey that sorts low-light image enhancement methods into traditional, deep learning, and hybrid categories, and it summarizes a handful of example pipelines. It is written as an entry point for beginners, but it contains no quantitative comparisons and does not achieve the comprehensive coverage promised in its title.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's central 'comprehensive' claim is unsupported: Sections II-IV describe roughly ten NTIRE challenge methods with no search protocol, inclusion criteria, or comparison tables, leaving standard low-light enhancement methods such as RetinexNet, Zero-DCE, and EnlightenGAN entirely absent.","rationale":"The reader's verdict is REJECT with high correctness risk, and my stress-test supports that. The load-bearing assumption is representativeness: without it, the survey cannot function as a reliable map or orientation guide. The paper gives no methodological basis for selection, and the content itself is internally inconsistent (wrong SID citation, dataset section misaligned with the stated contributions). A survey need not be exhaustive to be useful, but the title and abstract explicitly claim comprehensiveness; the gap between claim and content is the central failure. I found no reason to soften the verdict. The paper's summaries of the included NTIRE methods appear broadly plausible, so the issue is scope and rigor rather than accuracy of the described methods. A coverage audit would settle the matter definitively, but the available textual evidence already makes the representativeness assumption untenable.","tokens_in":10789,"tokens_out":3231,"duration_ms":31676,"concrete_test":"Compile a list of distinct low-light enhancement methods from recent peer-reviewed surveys, using the paper's own references [34] and [38] as seed sources, then count how many of those methods are cited or summarized in Sections II-IV. Separately verify whether the dataset section includes standard low-light enhancement datasets such as LOL, SID, and MIT-Adobe FiveK. If the coverage fraction of standard methods is below 20% and all described methods come from NTIRE challenges, the 'comprehensive' claim is refuted. Also check the Section IV.B citation: confirm that 'SID model [14]' should be [8].","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is to provide 'an extensive summary' and a three-way categorization of low-illumination image enhancement approaches. For that claim to hold, the methods selected in Sections II-IV must be representative of the field. They are not demonstrably so: every method described comes from NTIRE low-light/night photography challenges, and the paper never states a search protocol, inclusion criteria, or completeness guarantee. Classic and widely cited branches — retinex-based deep models, GAN-based enhancement, zero-reference methods, and illumination-map approaches — are absent from the body, even though the introduction mentions histogram equalization, gray-level transformations, and retinex models as traditional approaches. The datasets section (V) lists only color-constancy datasets (ColorChecker, NUS, Cube+, INTEL-TAU); it omits standard low-light enhancement datasets such as LOL and SID, undermining contribution 3. There is also an internal consistency error: Section IV.B refers to 'the pre-trained SID model [14]', but reference [14] is Gehler et al.'s color constancy paper; the SID paper is [8]. These omissions and errors mean a beginner using this text as a map would encounter a skewed, NTIRE-centric subset rather than a comprehensive survey. The individual method summaries are plausible, so this is a scope/representativeness failure, not a fabrication issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents itself as a comprehensive survey of low-illumination image enhancement. It organizes the field into traditional, learning-based, and hybrid approaches; describes roughly ten methods, mostly from NTIRE challenge entries; lists four color-constancy datasets; discusses open problems; and concludes with qualitative trade-offs. The paper contains no experiments, benchmarks, or quantitative comparisons, and its selection of methods and datasets is not justified by any stated search or inclusion protocol.","tokens_in":10954,"tokens_out":5163,"duration_ms":48889,"significance":"If the promised comprehensiveness were delivered, the survey would be a useful orientation guide for newcomers, and the proposed three-way taxonomy is reasonable. The individual method summaries are mostly faithful to the cited challenge papers, and the practice of attaching a limitations paragraph to each method is helpful. However, the main value of a survey—reliable coverage and representativeness—is not achieved: the method set is narrow and NTIRE-centric, the datasets are not low-light enhancement datasets, and no comparative basis is provided. The strengths are local; the central promise is not.","major_comments":[{"comment":"The title and abstract promise an 'extensive summary' and a 'comprehensive' review, and Section I states that the paper serves as a practical guide for beginners. The body does not support this: Sections II-IV describe roughly ten methods, all drawn from NTIRE challenge reports, with no search protocol, inclusion criteria, or completeness guarantee. Widely cited branches of the field—retinex-based deep models, GAN-based enhancement such as EnlightenGAN, zero-reference methods such as Zero-DCE, and illumination-map approaches—are absent from the body even though the introduction mentions retinex, histogram equalization, and gray-level transformations as traditional approaches. This representativeness gap is load-bearing because the paper's stated contribution is to 'review and categorize low-illumination image enhancement approaches.'","section":"Abstract and Sections I-IV"},{"comment":"Section V, 'Datasets,' covers only color-constancy datasets (ColorChecker, NUS 8-Camera, Cube+, INTEL-TAU) and contains no standard low-illumination enhancement dataset such as LOL or SID. This directly contradicts contribution 3, which promises 'an overview of various datasets for low-illumination image enhancement and related tasks,' and it is internally inconsistent with the body: SID is used in the method descriptions of Sections III.C and IV.B but never appears in the dataset section.","section":"Section V and Contribution 3"},{"comment":"The survey lacks any quantitative comparison or evaluation framework: there are no benchmark tables, metric definitions, or reproducible results, and the 'advantages and limitations' statements are qualitative assertions without supporting evidence. Because contribution 2 is explicitly to discuss limitations of each approach, the absence of any comparative evaluation leaves the reader unable to judge the claimed trade-offs among traditional, learning-based, and hybrid methods.","section":"Sections II-VII"}],"minor_comments":[{"comment":"The text says 'the pre-trained SID model [14]', but reference [14] is Gehler et al.'s color constancy paper; the SID paper is reference [8] (Chen et al., Learning to See in the Dark).","section":"Section IV.B"},{"comment":"'Quai-WB' appears to be a typo for 'Quasi-WB' (the method of reference [5]).","section":"Section III.B"},{"comment":"'LEARNING-BASED APPRAOCHES' contains a typo ('APPRAOCHES'), and Figure 1's 'Taxanomy' should be 'Taxonomy'.","section":"Section III heading and Figure 1"},{"comment":"The first sentence, 'Ordinance Normalization in Batch The ResNet (BINResNet) model is utilized to improve images,' is garbled and should be rewritten to describe the Batch-Instance Normalization ResNet.","section":"Section III.E"},{"comment":"The first bullet contains an extra comma after the colon in 'Noise Reduction and Detail retention: , In low-illumination...' and should be corrected.","section":"Section VI"}],"recommendation":"reject","confidential_remarks":"The manuscript is a plausible but narrow summary of NTIRE challenge methods. The mismatch between the title/abstract and the content is substantial; to make the central claim defensible the authors would need either to expand coverage to the broader low-light enhancement literature with a stated search protocol or to rewrite the paper as a focused review of NTIRE challenge methods. I do not see a path to acceptance under the current claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nHere’s the quick version: this is a catalog of about a dozen NTIRE low-light night photography challenge pipelines, wrapped in a title that promises a comprehensive survey. The body does a decent job of summarizing the individual challenge entries in plain language, but the load-bearing comprehensiveness claim does not hold up.\n\nThe three-way split (traditional, learning-based, hybrid) is standard—the paper’s own references include earlier surveys with the same organizing scheme. What the paper does well is give compact paraphrases of the specific methods it covers: the two-stage traditional pipeline with LCC and BM3D, the three-stage cascaded framework, U-ISP, BINResNet, and the hybrid pretrained-SID approach. Those summaries are broadly faithful to the cited sources, and a reader who wants a five-minute description of those particular NTIRE contributions will get it here.\n\nThe soft spots are substantial. The abstract and title claim an 'extensive summary' and 'comprehensive' coverage, but the body only describes roughly ten methods, all from NTIRE challenges. There is no search protocol, no inclusion criteria, no comparison tables, no quantitative evaluation, and no discussion of evaluation metrics. Classic branches of low-light enhancement—Retinex-based deep models, zero-reference methods, GAN-based enhancement—are missing entirely. The datasets section only lists color constancy datasets (ColorChecker, NUS, Cube+, INTEL-TAU), omitting LOL and SID, which is a strange choice for a survey on low-light enhancement. And there is an easy-to-catch citation error: Section IV.B refers to the 'pre-trained SID model [14]', but [14] is Gehler et al.’s color constancy paper; the SID paper is [8]. These are not cosmetic problems, because they directly undermine the tutorial value of the survey.\n\nNo equations, no code, no machine-checked proofs to weigh; this is entirely a descriptive review. The closest thing to credit is that the individual method summaries seem accurate.\n\nMy bottom line: the paper is for someone who wants a very quick look at the NTIRE challenge solutions specifically, not for anyone seeking a map of the low-light enhancement field. Because the central claim is overstated and the coverage is demonstrably skewed, I would not accept this for peer review, and I would not cite it as a survey. It could be reshaped into a modest 'selected NTIRE pipeline overview,' but not in its current form.\n\nBest,\n[You]","headline":"A well-meaning compilation of NTIRE pipeline summaries that fails the 'comprehensive' claim in its title; desk-reject as a survey, though the individual method descriptions are mostly accurate.","tokens_in":11553,"tokens_out":6841,"would_cite":false,"duration_ms":56722,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey sorts low-illumination image enhancement into three method families: traditional, learning-based, and hybrid.","keywords":["low-illumination image enhancement","image signal processing","survey","deep learning","traditional methods","hybrid methods","image denoising","white balance"],"falsifier":"Counting all peer-reviewed low-light image enhancement methods in a standard bibliographic database and comparing the result with the number described here would settle whether the survey is comprehensive; if entire families such as diffusion-based or generative-adversarial methods are absent, the paper's central claim fails. An absence of any quantitative benchmark table in the survey also leaves the claim that learning-based methods dominate untestable as presented.","tokens_in":10481,"feed_emoji":"🌙","tokens_out":7002,"duration_ms":61758,"temperature":0.7,"pith_summary":"This paper sets out to provide a comprehensive survey of image signal processing (ISP) methods for enhancing images captured in low light. The author's central claim is that these methods fall into three families: traditional approaches that adjust brightness, contrast, and noise; learning-based approaches that use convolutional neural networks to map low-light inputs to improved images; and hybrid approaches that combine both. The paper also lists datasets commonly used in the field and identifies open research problems such as noise-detail tradeoffs, extreme conditions, real-time processing, and dataset diversity. If the survey is right, a newcomer can use it as an orientation guide to the main approaches and their tradeoffs.","feed_headline":"Survey sorts low-light image enhancement into three families","feed_subtitle":"Traditional, learning-based, and hybrid approaches mapped, with datasets and open problems for newcomers.","key_machinery":"The organizing object of the survey is the camera image signal processing pipeline—the sequence of demosaicing, white balance, color correction, tone mapping, and denoising that turns raw sensor data into a displayable image. The paper uses this pipeline as the backbone for its taxonomy: traditional methods operate on individual pipeline stages with hand-crafted operators, learning-based methods replace stages or the whole pipeline with neural networks, and hybrid methods keep traditional front-end processing while adding learned tone mapping or enhancement. The taxonomy is what carries the argument that the field can be understood as variations on a common pipeline.","core_discovery":"The core discovery of this review is a three-way taxonomy of low-illumination image enhancement. Traditional methods, illustrated by two-stage pipelines and low-light rendering with separate noise suppression, are computationally cheap but struggle with complex scenes and color casts. Learning-based methods, demonstrated through cascaded frameworks, specialized ISP pipelines, transformer-based denoisers, and generative models, currently lead the field because they can suppress noise while preserving detail, at the cost of large training datasets and compute. Hybrid approaches stitch traditional ISP stages together with learned tone mapping or enhancement networks and are presented as a promising middle ground. The paper further compiles a set of public datasets (color-checker images, multi-camera color constancy sets, and paired RAW/sRGB low-light data) and a list of open challenges.","pith_inferences":["The methods described appear to be drawn mainly from one recent challenge rather than from a systematic literature search; a reader should not assume that omitted methods are less significant.","A quantitative comparison table would be needed to actually verify the claimed tradeoffs between speed, quality, and data requirements; the survey does not provide one.","The ISP-pipeline framing suggests that low-light enhancement is best decomposed into subproblems (denoising, white balance, tone mapping), a view that could be tested by measuring whether modular systems outperform end-to-end models on out-of-distribution scenes.","One testable extension is to apply the same taxonomy to a broader corpus of published methods and check whether the three families indeed cluster by computational cost and performance."],"forward_implications":["A beginner can use the three-family taxonomy to decide which class of methods to study first: traditional for quick baselines, learning-based for state-of-the-art quality, hybrid for a balance.","The paper's reading implies that deep learning will continue to dominate low-light enhancement, with the main bottlenecks being training data volume and inference speed.","The listed open problems—noise-detail balance, extreme dynamic range, real-time processing, and dataset diversity—define concrete research targets for the field.","Hybrid methods, which combine cheap classical preprocessing with learned tone mapping, are presented as a practical route to production systems.","The datasets reviewed provide starting points for training and benchmarking, though the paper notes that broader and more diverse collections are still needed."],"supporting_citations":[{"why":"Supplies the BM3D denoising algorithm used in the two-stage traditional pipeline.","marker":"[11]"},{"why":"Provides the trilateral weighted sparse coding denoiser used in the low-light rendering and noise suppression approach.","marker":"[39]"},{"why":"Contributes the multi-exposure fusion method used for low-light rendering before denoising.","marker":"[40]"},{"why":"Provides a convolutional encoder-decoder backbone used for RAW denoising in the three-stage cascaded framework.","marker":"[32]"},{"why":"Supplies the learned Bayer-to-RGB mapping network used in the cascaded framework.","marker":"[19]"},{"why":"Provides a transformer-based image restoration model used for sRGB denoising in the mixed white-balance pipeline.","marker":"[27]"},{"why":"Supplies a paired RAW/sRGB low-light dataset used to train the RAW-based enhancement approach.","marker":"[8]"},{"why":"Provides a large illumination estimation dataset used for training white balance and color constancy models.","marker":"[13]"},{"why":"Supplies a learned tone-mapping network used in the traditional ISP plus CNN hybrid approach.","marker":"[15]"}],"fun_headline_variants":["Survey maps low-light enhancement: classic, deep, hybrid","Deep learning leads low-light image enhancement, survey finds","Hybrid methods bridge classic and deep low-light enhancement","Survey catalogs low-light enhancement: methods, datasets, gaps","Low-light image enhancement: three families of approaches"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's value as a map depends on the roughly ten methods it describes being representative of the whole low-illumination enhancement literature, but the paper does not state a search protocol or inclusion criteria, so the sample may not span the field.","fun_headline_variants_meta":{"raw":{"variants":["Survey maps low-light enhancement: classic, deep, hybrid","Deep learning leads low-light image enhancement, survey finds","Hybrid methods bridge classic and deep low-light enhancement","Survey catalogs low-light enhancement: methods, datasets, gaps","Low-light image enhancement: three families of approaches"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000303,"raw_usage":{"total_tokens":1748,"prompt_tokens":956,"completion_tokens":792,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":724}},"tokens_in":572,"tokens_out":792,"duration_ms":7291,"temperature":1.0,"reasoning_tokens":724,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:07:17.567603+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Counting all peer-reviewed low-light image enhancement methods in a standard bibliographic database and comparing the result with the number described here would settle whether the survey is comprehensive; if entire families such as diffusion-based or generative-adversarial methods are absent, the paper's central claim fails. An absence of any quantitative benchmark table in the survey also leaves the claim that learning-based methods dominate untestable as presented.","supporting_citations":[{"cited_title":"Image denoising by sparse 3-d transform-domain collaborative filtering","cited_arxiv_id":null,"evidence_quote":"Supplies the BM3D denoising algorithm used in the two-stage traditional pipeline."},{"cited_title":"A trilateral weighted sparse coding scheme for real-world image denoising, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the trilateral weighted sparse coding denoiser used in the low-light rendering and noise suppression approach."},{"cited_title":"A bio-inspired multi-exposure fusion framework for low-light image enhancement, 2017","cited_arxiv_id":null,"evidence_quote":"Contributes the multi-exposure fusion method used for low-light rendering before denoising."},{"cited_title":"Aim 2020 challenge on learned image signal processing pipeline, 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the learned Bayer-to-RGB mapping network used in the cascaded framework."},{"cited_title":"Learning to see in the dark, 2018","cited_arxiv_id":null,"evidence_quote":"Supplies a paired RAW/sRGB low-light dataset used to train the RAW-based enhancement approach."},{"cited_title":"The cube++ illumination estimation dataset","cited_arxiv_id":null,"evidence_quote":"Provides a large illumination estimation dataset used for training white balance and color constancy models."},{"cited_title":"Barron, Samuel W","cited_arxiv_id":null,"evidence_quote":"Supplies a learned tone-mapping network used in the traditional ISP plus CNN hybrid approach."}],"review_version":1}