{"id":"c9966e00-e77a-4765-8845-5a23fb9d6b1b","arxiv_id":"2501.11310","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured survey of vision-based industrial anomaly detection covering acquisition, preprocessing, learning, evaluation, datasets, and future directions; no new result.","lead":"Industrial anomaly detection uses cameras and machine learning to spot defects on factory products. This review organizes the field by pipeline stage, datasets, and open challenges, but adds no new experiments.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Undefined 'low-ranked journal' exclusion and missing included-paper list make the survey's literature selection and Fig. 5 trend unreproducible, weakening the central coverage claim.","rationale":"The central claim of this survey is that it provides an up-to-date, structured overview of vision-based industrial anomaly detection with attention to pipeline components overlooked by prior surveys. For a survey, the value of that claim depends on the representativeness and transparency of the literature selection. The paper specifies a search string and databases but leaves the 'low-ranked journal papers' exclusion undefined and does not publish the final list of included papers. This is the weakest link in the argument because every downstream synthesis—the trend in Fig. 5, the comparison table of existing surveys, the pipeline tables, and the dataset summaries—rests on an unverifiable corpus. The reader's weakest_assumption identified exactly this representativeness risk, and I agree. I also examined the Table VI MVTec AD inconsistency highlighted by the reader: the text's 1258 defective images refer to the testing bank, while the table's 1725 defective images match the dataset total, so the apparent conflict may be a labeling ambiguity rather than a factual error. Therefore I do not elevate that to the primary concern. A concrete replication test—rerunning the search, comparing against a gold-standard list, and recomputing the trend under explicit ranking definitions—would settle whether the selection bias concern actually lands. If the test shows the corpus is broadly representative and the trend is stable, the survey's central claim holds; if not, the survey would need to disclose its full selection protocol and either correct its coverage statements or narrow its claims. The correctable nature of the flaw supports the reader's CONDITIONAL verdict rather than a stronger rejection.","tokens_in":934,"tokens_out":865,"duration_ms":52415,"concrete_test":"Replicate the Section I-B search on IEEE Xplore, Scopus, Springer, and Taylor & Francis for 2019 through June 2024 using the stated query. Apply 'low-ranked journal' exclusion in two well-defined ways (e.g., no exclusion; exclude journals below JCR Q2), and compare the deduplicated title set against the paper's reference list and against a manually assembled gold-standard list of influential IAD works (e.g., MVTec AD, PatchCore, WinCLIP, PaDiM, VisA, Real-IAD, and their primary source papers). Recompute the year-by-year publication counts and the reported 380% growth figure. If the paper's set omits a substantial fraction of the gold-standard works, or if the trend changes materially between the two ranking definitions, the survey's coverage and trend claims are not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section I-B states that relevant literature was collected via a keyword search on four databases, with 'low-ranked journal papers were excluded from the final selection,' but no ranking criterion is defined and no list of included papers is provided. The reference list mixes journal articles, conference papers, and arXiv preprints, so the practical effect of the exclusion rule cannot be inferred. Without a reproducible inclusion/exclusion procedure, the yearly counts in Fig. 5, the '380% increase from 2019 to 2023' claim, and the coverage implied by Tables I, II, III, and V cannot be independently checked. If the selection is biased by venue, database, or the authors' unstated ranking rule, the survey's stated focus on pipeline components that are 'overlooked in existing surveys'—data acquisition, preprocessing, learning mechanisms, and evaluation—could reflect the chosen corpus rather than the actual state of the field. The Table VI discrepancy noted by the reader (MVTec AD: 1258 test-set defects in the text vs 1725 defective images in the table) is less decisive, because the table column appears to report total defective images while the text reports only the testing subset; the selection problem is the more load-bearing concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of vision-based industrial anomaly detection (IAD) that reviews literature published from 2019 onward. It organizes the field around four pipeline components that the authors claim are under-covered in prior surveys: data acquisition, preprocessing, learning mechanisms, and evaluation. It also discusses scientific and industrial challenges, summarizes popular datasets, and outlines future directions such as explainable AI and vision-language models. The paper positions itself as a reference for researchers and practitioners entering or working in industrial visual inspection.","tokens_in":27425,"tokens_out":3722,"duration_ms":38389,"significance":"If its coverage is reliable, this survey would be a useful entry point for the IAD community: it consolidates a wide range of recent references, gives a clear pipeline-oriented structure, and includes practical industrial considerations such as edge processing and real-time constraints. The paper explicitly names the overlooked pipeline components in its contributions and provides a dataset table that practitioners often need. A notable strength is the breadth of collected application scenarios, from PCB and aero-engine inspection to agricultural and bridge defect detection, which gives the survey interdisciplinary reach. However, the survey's utility as a reference depends on the reproducibility of its literature selection and on the accuracy of its dataset summaries, both of which are currently problematic.","major_comments":[{"comment":"The literature selection procedure is not reproducible, which undermines the central coverage claim. Section I-B states that 'low-ranked journal papers were excluded from the final selection' but does not define the ranking criterion (e.g., journal quartile, impact factor, or a specific list). No complete list of included papers, no PRISMA-style screening flow, and no counts of retrieved, screened, and excluded records are given. Consequently, the quantitative claim of 'nearly 380%' growth from 2019 to 2023 and the yearly counts in Figure 5 cannot be independently verified. The authors should specify the ranking rule, provide the full list of included studies (e.g., in a supplementary file), and report screening statistics so that the trend and the coverage claims can be audited.","section":"§I-B and Figure 5"},{"comment":"Table VI contains an internal inconsistency with Section V for the MVTec AD dataset. The text states that the testing bank includes 1258 defect images, while Table VI lists 'Defective images' as 1725. The number 1725 is actually the total test-set size (1258 defective plus 467 non-defective), so the column label is misleading. Similar issues may affect other rows: for KSDD2, the description in Table VI says the training set has '246 and 2085 images respectively' and the testing set has '110 and 894 images respectively,' yet the table's totals are 2979 non-defective and 356 defective, which do not correspond to those breakdowns. Because the dataset summary is one of the paper's stated contributions, each row must be reconciled with the original dataset papers and the sources should be cited per row.","section":"§V and Table VI"}],"minor_comments":[{"comment":"The statement that self-supervised learning 'does not require a loss function' is incorrect; self-supervised methods typically optimize pretext-task losses (e.g., contrastive or reconstruction losses).","section":"§I-A2"},{"comment":"The description of GANs includes 'The algorithm generates region proposals in an image and uses CNN to classify each proposal,' which describes an R-CNN-style detector rather than a generative adversarial network; this should be corrected to avoid confusing readers.","section":"§IV-C, GAN paragraph"},{"comment":"The text defines a low false positive rate as 'dealing with the number of defective instances being incorrectly classified as non-defective'; this is actually a false negative, not a false positive. The definition should be fixed.","section":"§III-A, small defects bullet"},{"comment":"Figure 7 refers to a 'highly cited survey' identified with a red flag, but no citation threshold or basis for that designation is given; please clarify or remove the qualifier.","section":"§II, Figure 7 caption"},{"comment":"Several passages in the learning-mechanisms section read as textbook definitions (e.g., SVM, KNN, decision trees) without tying them to the surveyed IAD literature; adding a summary table linking each algorithm to representative industrial applications would strengthen the survey's utility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The two major issues—selection reproducibility and dataset-table accuracy—are both fixable within the manuscript's scope, but they are load-bearing for a survey that claims to provide a reliable structured overview. I would not reject; the pipeline framing and breadth are valuable. The authors should be asked to supply a complete list of included papers and a clear exclusion rule, and to re-verify every row of Table VI against the primary sources. I also noticed that the reference list includes several arXiv preprints and that some cited statistics are not from the originating papers; the editor may wish to ask for a careful reference-by-reference check."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a survey, not a research paper, and it should be evaluated as one. Its main value is organizational—framing IAD through the full pipeline (data acquisition, preprocessing, learning, evaluation) rather than just models. That angle is genuinely less covered in earlier surveys, and the dataset roundup (MVTec, KSDD2, VisA, AeBAD, Real-IAD, VAD) is current enough to be handy.\n\nWhat it does well: the pipeline framing is coherent, the challenges section is sensible, and the comparison of prior surveys in Table I is fair. The paper doesn't overclaim: it says it organizes and reviews, and that's what it does.\n\nSoft spots, in order of importance. First, the literature-selection procedure is too vague to reproduce. Section I-B says 'low-ranked journal papers were excluded' but never defines 'low-ranked,' and the final included-paper list isn't given. That matters because Fig. 5 and the '380% increase' claim depend entirely on the search corpus. If the selection is skewed by venue or the authors' unstated ranking rule, the pipeline coverage claim could reflect the corpus rather than the field. This is fixable but needs to be addressed. Second, Table VI has a concrete error: MVTec AD is listed with 1,725 defective images, but the text correctly says 1,258 defect images in the testing bank; 1,725 is the total test-set size. That's a minor data error, but in a survey whose function is reliable summarization, it matters.\n\nThere are also some copyediting issues (typos, odd phrasing), but nothing that affects the substance.\n\nBottom line: for a practitioner or new researcher wanting a structured entry point to industrial anomaly detection, this survey is useful. It's not a major scientific advance, and the flaws are correctable rather than load-bearing. I'd send it to peer review, but with a clear request to document the inclusion/exclusion criteria, list the included papers, and fix the dataset table. If I were working in IAD, I might cite it for the pipeline framing, but not before the transparency issues are fixed.","headline":"Useful organizational survey of industrial anomaly detection, but the opaque literature selection and a dataset table error keep it from being a reliable reference as-is.","tokens_in":28019,"tokens_out":3006,"would_cite":false,"duration_ms":28908,"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 survey organizes vision-based industrial anomaly detection around four pipeline stages — data acquisition, preprocessing, learning mechanisms, evaluation — and uses them to frame the field's challenges, datasets, and future directions.","keywords":["Anomaly inspection","computer vision","deep learning","industrial defect detection","neural networks","quality control","vision sensors","industrial datasets"],"falsifier":"Re-run the paper's stated keyword search on the four scientific databases it lists, for 2019 through June 2024, and compare the retrieved set with the papers summarized in the survey's tables; if a large share of relevant vision-based IAD papers is missing, or if the claimed roughly 380% growth in publications does not reproduce, the coverage claim is falsified. A second check is to verify whether each earlier survey the paper lists as limited — for example, restricted to surface defects, unsupervised methods, or single domains — actually has that limitation.","tokens_in":27041,"feed_emoji":"🏭","tokens_out":6656,"duration_ms":60639,"temperature":0.7,"pith_summary":"Vision-based industrial anomaly detection (IAD) has shifted from human inspection to automated systems that photograph manufactured parts and classify defects. This survey, covering work published since 2019, claims that earlier reviews concentrate on detection models alone and neglect the surrounding pipeline. It therefore builds its review around four pipeline stages — data acquisition, preprocessing, learning mechanisms, and evaluation — and also catalogs popular industrial datasets and open scientific and industrial challenges. A reader who wants to enter the field or situate a particular method gets a structured map of where methods, datasets, and evaluation practices sit relative to each other.","feed_headline":"Survey maps industrial defect detection from camera to verdict","feed_subtitle":"Reviews vision-based studies since 2019 and covers the pipeline stages, datasets, and challenges other surveys skip.","key_machinery":"The carrying structure is the vision-based IAD pipeline itself, defined as four stages: data acquisition, data preprocessing, learning mechanisms, and evaluation. The survey uses this pipeline as a taxonomy: every reviewed method, dataset, and reported score is placed in one of its stages, so the pipeline does the work of showing which parts of the field are well covered and which are thin. Secondary organizing devices are the learning-paradigm split (supervised, unsupervised, semi-supervised) and the dataset comparison table that ties performance claims to specific benchmarks.","core_discovery":"On the paper's own terms, the central discovery is organizational: the published literature on vision-based IAD is best understood not as a stack of competing models but as a four-stage pipeline. Data acquisition (line-scan, multi-view, omnidirectional, and infrared thermographic imaging) sets the limits of what can be seen; preprocessing (grayscale conversion, sharpening, rotation, resizing, thresholding, wavelet transform) prepares raw images for learning; learning mechanisms (from SVM and KNN to CNN, GAN, YOLO, teacher-student, few-shot, and zero-shot vision-language models) carry the classification; and evaluation metrics (precision, accuracy, recall, specificity, F1, AUROC, IOU) decide what counts as success. The survey further claims that this pipeline view exposes challenges other reviews underplay — real-time inspection, small and imbalanced datasets, annotation cost, data quality, and system integration — and that recent benchmarks such as Real-IAD show current models saturating on older datasets.","pith_inferences":["The pipeline framing implies that evaluation practice is a weak link: many summarized papers report only one or two metrics on a single dataset, so cross-paper comparisons are fragile; a testable extension would be a standardized evaluation protocol spanning all four pipeline stages.","Emphasis on data acquisition suggests that IAD progress may be bottlenecked more by sensing and annotation than by model architecture; a plausible extension is that cheaper multi-view and infrared imaging plus generative synthesis of labeled defects could advance the field faster than new loss functions.","The saturation of MVTec AD and KSDD implied by recent results suggests that benchmarks with domain shift and logical anomalies (AeBAD, MVTec LOCO AD, Real-IAD) will become the new standard, and re-ranking existing methods on these harder benchmarks would be a direct test.","The vision-language future direction implies that zero-shot detection could drastically cut annotation cost; one could test whether LVLM-based detectors keep their advantage on rare, tail-case defects rather than frequent patterns."],"forward_implications":["Researchers entering the field can use the four-stage pipeline as a checklist for what a complete IAD system must include, rather than focusing only on the model.","The survey identifies recurring industrial obstacles — real-time detection, data imbalance, annotation cost, data quality, and system integration — and collects candidate remedies such as edge processing, data augmentation, generative synthesis, and few-shot learning.","The reported evaluation scores across applications make plain that no single metric tells the whole story; the metric formulas give practitioners the vocabulary to compare methods fairly.","The dataset summary shows a trajectory from small single-defect datasets like KSDD to large multi-view benchmarks like Real-IAD, where state-of-the-art image AUROC drops from 97.9% on MVTec AD to 85% on Real-IAD, indicating that older benchmarks are nearing saturation.","Future directions named by the paper — explainable AI and large vision-language models — point toward inspection systems that can justify their decisions and work with little or no labeled defect data."],"supporting_citations":[{"why":"Prior survey of ML/DL in operation management; the paper contrasts its own pipeline scope and post-2019 coverage against this broader view.","marker":"[12]"},{"why":"Survey of unsupervised anomaly localization; the paper positions its inclusion of supervised and semi-supervised methods as the gap this review fills.","marker":"[21]"},{"why":"Deep image anomaly detection survey; used as a comparison point that lacks industrial sector and application context.","marker":"[26]"},{"why":"Recent DL surface-defect survey; the paper's claim to cover non-surface defects and the full pipeline is built against this work.","marker":"[27]"},{"why":"KSDD dataset of electrical commutator defects; a small single-defect benchmark whose saturation motivates newer datasets.","marker":"[96]"},{"why":"MVTec AD, the default unsupervised benchmark; most reviewed methods are evaluated on it, so it anchors the dataset and evaluation discussion.","marker":"[113]"},{"why":"KSDD2 dataset with multiple defect types; shows the move beyond KSDD's saturation and supports the dataset-trajectory claim.","marker":"[116]"},{"why":"AeBAD aero-engine blade dataset with domain shift; underpins the aerospace application and domain-shift discussion.","marker":"[122]"},{"why":"Real-IAD large multi-view dataset; the AUROC drop to 85% on it supports the claim that current benchmarks are nearing saturation.","marker":"[123]"}],"fun_headline_variants":["New survey dissects industrial anomaly detection pipeline gaps","Industrial anomaly detection review zooms into overlooked pipeline steps","Four-stage pipeline frames industrial defect detection review","Vision-based anomaly detection survey covers full inspection pipeline","Review unpacks four-stage pipeline for industrial anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's usefulness depends on the selection of papers being representative of the field; the authors excluded 'low-ranked journal papers' without defining the ranking criterion or publishing the full list of included papers, so a biased or incomplete selection would make the reported trends and coverage claims unreliable.","fun_headline_variants_meta":{"raw":{"variants":["New survey dissects industrial anomaly detection pipeline gaps","Industrial anomaly detection review zooms into overlooked pipeline steps","Four-stage pipeline frames industrial defect detection review","Vision-based anomaly detection survey covers full inspection pipeline","Review unpacks four-stage pipeline for industrial anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000627,"raw_usage":{"total_tokens":2905,"prompt_tokens":958,"completion_tokens":1947,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1875}},"tokens_in":574,"tokens_out":1947,"duration_ms":14437,"temperature":1.0,"reasoning_tokens":1875,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:24:10.867575+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the paper's stated keyword search on the four scientific databases it lists, for 2019 through June 2024, and compare the retrieved set with the papers summarized in the survey's tables; if a large share of relevant vision-based IAD papers is missing, or if the claimed roughly 380% growth in publications does not reproduce, the coverage claim is falsified. A second check is to verify whether each earlier survey the paper lists as limited — for example, restricted to surface defects, unsupervised methods, or single domains — actually has that limitation.","supporting_citations":[{"cited_title":"Segmentation-based deep-learning approach for surface-defect detection,","cited_arxiv_id":null,"evidence_quote":"KSDD dataset of electrical commutator defects; a small single-defect benchmark whose saturation motivates newer datasets."},{"cited_title":"Mixed supervision for surface- defect detection: from weakly to fully supervised learning,","cited_arxiv_id":null,"evidence_quote":"KSDD2 dataset with multiple defect types; shows the move beyond KSDD's saturation and supports the dataset-trajectory claim."},{"cited_title":"Industrial anomaly detection with domain shift: A real-world dataset and masked multi-scale reconstruction,","cited_arxiv_id":null,"evidence_quote":"AeBAD aero-engine blade dataset with domain shift; underpins the aerospace application and domain-shift discussion."}],"review_version":1}