Frequency‑Enhanced Dual‑Layer Anomaly Synthesis for Real‑Time Industrial Surface Inspection

Frequency‑Enhanced Dual‑Layer Anomaly Synthesis for Real‑Time Industrial Surface Inspection | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Frequency‑Enhanced Dual‑Layer Anomaly Synthesis for Real‑Time Industrial Surface Inspection ChihYuan Chen, YuHung Chiang, ChingHua Hung, ChunWei W. Wu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7240661/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Micro-scale linear guideways increasingly rely on face-seal gaskets whose defects are minute and low-contrast; such defects must be inspected within 20 s, a requirement that increases miss rates. The high cost of pixel-level labels limits the practicality of fully supervised convolutional neural network (CNN)-based approaches. To address these constraints, we propose GLASS-FFT-SA, a framework that fuses a dual-layer anomaly synthesis strategy with frequency domain convolution and a lightweight spectralattention (SA) block that selectively amplifies highfrequency defect cues. Local Anomaly Synthesis inserts defect textures into normal images to create strong anomalies, whereas global anomaly synthesis performs gradient ascent on the feature manifold to generate subtle, near-boundary anomalies; together, they furnish abundant, diverse training data. Replacing the large 7 × 7 and 5 × 5 kernels in a ResNet-34 backbone with FFT-based convolutions reduces the computational complexity of those layers and reduces inference latency by approximately 30%, enabling near-real-time operation. To avoid running the pixel head on every frame, we introduce a gated crossattention mechanism (GCAM) that activates the pixel branch only when the image head’s anomaly score ŷ exceeds a learnable hardsigmoid gate. Trained on 10,000 normal images and 8,000 synthetically generated anomalies, GLASS-FFT-SA achieved an image-level AUROC of 0.99, a pixel-level AUROC of 0.97, an AUPRO of 0.95, and a throughput of 38 FPS on RTX 3090—matching the precision of original GLASS while operating at approximately 40% higher speed. It also sustains AUROCs above 0.93 across variations in product type, illumination, and resolution, outperforming conventional CNNs, ResNet-34-FFT-SA, and PatchCore. These results suggest that combining spectrum-efficient convolutions with tailored anomaly synthesis can deliver both accuracy and throughput for fine-grained defect inspection. Seal pad inspection Anomaly detection GLASS-FFT-SA Synthetic anomalies Frequency-domain convolution Full Text Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7240661","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":510994356,"identity":"a34281d2-512a-4296-90d7-4825e003d527","order_by":0,"name":"ChihYuan Chen","email":"","orcid":"","institution":"National Yang Ming Chiao Tung University","correspondingAuthor":false,"prefix":"","firstName":"ChihYuan","middleName":"","lastName":"Chen","suffix":""},{"id":510994358,"identity":"311e48bd-c6b5-439e-874a-506d36d811e4","order_by":1,"name":"YuHung Chiang","email":"","orcid":"","institution":"National University of Kaohsiung","correspondingAuthor":false,"prefix":"","firstName":"YuHung","middleName":"","lastName":"Chiang","suffix":""},{"id":510994359,"identity":"042b324d-e334-44a5-9919-e9f2cf91554f","order_by":2,"name":"ChingHua Hung","email":"","orcid":"","institution":"National Yang Ming Chiao Tung University","correspondingAuthor":false,"prefix":"","firstName":"ChingHua","middleName":"","lastName":"Hung","suffix":""},{"id":510994360,"identity":"602d1cc6-a1d0-46f9-be00-ba57ed3fa051","order_by":3,"name":"ChunWei W. Wu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6klEQVRIiWNgGAWjYFAC5gMHEgwk6vnZmw8AeRIyRGhhS3zwocImQbLnWAJICw8RWniMDWecSUswuOFjAOYS1MDff8BMmrftcB7DDZ7Pr27UWPAwsB8+ugGfFokbCWkgLcWMs3u3WeccAzqMJy3tBl5rbjAcA2lhbJY5u804hw2oRYLHDK8W+fMH28Ba2iRynhnn/CNCi8GBZGaQ9xN7JHKYH+e2EaHF8EYaIyiQjSV4jpkx5/ZJ8LAR8ovc+fMfQFEpZ3+8+fHnnG91cvzsh4/h9z4SYJMAk8QqBwHmD6SoHgWjYBSMgpEDABUgTUWYsxMFAAAAAElFTkSuQmCC","orcid":"","institution":"National Sun Yat-sen University","correspondingAuthor":true,"prefix":"","firstName":"ChunWei","middleName":"W.","lastName":"Wu","suffix":""}],"badges":[],"createdAt":"2025-07-29 08:08:29","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7240661/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7240661/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92191945,"identity":"740cce7b-8932-4537-be7f-6d58fb22c160","added_by":"auto","created_at":"2025-09-25 15:14:09","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1240865,"visible":true,"origin":"","legend":"","description":"","filename":"20250729Manuscriptanonymousvfianl.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7240661/v1_covered_343ac2d0-4a50-4d59-9fe6-f482a59a52b2.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Frequency‑Enhanced Dual‑Layer Anomaly Synthesis for Real‑Time Industrial Surface Inspection","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Seal pad inspection, Anomaly detection, GLASS-FFT-SA, Synthetic anomalies, Frequency-domain convolution","lastPublishedDoi":"10.21203/rs.3.rs-7240661/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7240661/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eMicro-scale linear guideways increasingly rely on face-seal gaskets whose defects are minute and low-contrast; such defects must be inspected within 20 s, a requirement that increases miss rates. The high cost of pixel-level labels limits the practicality of fully supervised convolutional neural network (CNN)-based approaches. To address these constraints, we propose GLASS-FFT-SA, a framework that fuses a dual-layer anomaly synthesis strategy with frequency domain convolution and a lightweight spectralattention (SA) block that selectively amplifies highfrequency defect cues. Local Anomaly Synthesis inserts defect textures into normal images to create strong anomalies, whereas global anomaly synthesis performs gradient ascent on the feature manifold to generate subtle, near-boundary anomalies; together, they furnish abundant, diverse training data. Replacing the large 7 \u0026times; 7 and 5 \u0026times; 5 kernels in a ResNet-34 backbone with FFT-based convolutions reduces the computational complexity of those layers and reduces inference latency by approximately 30%, enabling near-real-time operation. To avoid running the pixel head on every frame, we introduce a gated crossattention mechanism (GCAM) that activates the pixel branch only when the image head\u0026rsquo;s anomaly score ŷ exceeds a learnable hardsigmoid gate. Trained on 10,000 normal images and 8,000 synthetically generated anomalies, GLASS-FFT-SA achieved an image-level AUROC of 0.99, a pixel-level AUROC of 0.97, an AUPRO of 0.95, and a throughput of 38 FPS on RTX 3090\u0026mdash;matching the precision of original GLASS while operating at approximately 40% higher speed. It also sustains AUROCs above 0.93 across variations in product type, illumination, and resolution, outperforming conventional CNNs, ResNet-34-FFT-SA, and PatchCore. These results suggest that combining spectrum-efficient convolutions with tailored anomaly synthesis can deliver both accuracy and throughput for fine-grained defect inspection.\u003c/p\u003e","manuscriptTitle":"Frequency‑Enhanced Dual‑Layer Anomaly Synthesis for Real‑Time Industrial Surface Inspection","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-09 16:55:23","doi":"10.21203/rs.3.rs-7240661/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"be6380c2-d376-42b7-920b-fd695a93a14f","owner":[],"postedDate":"September 9th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-25T15:12:19+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-09 16:55:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7240661","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7240661","identity":"rs-7240661","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}