Anomaly-Focused Augmentation Method for Industrial Visual Inspection

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F61989100%3A27240%2F25%3A10257712" target="_blank" >RIV/61989100:27240/25:10257712 - isvavai.cz</a>

Výsledek na webu

<a href="https://ieeexplore.ieee.org/document/11003961/authors#authors" target="_blank" >https://ieeexplore.ieee.org/document/11003961/authors#authors</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1109/ACCESS.2025.3570075" target="_blank" >10.1109/ACCESS.2025.3570075</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Anomaly-Focused Augmentation Method for Industrial Visual Inspection

Popis výsledku v původním jazyce

In industrial inspection, the detection of surface defects—such as scratches, dents, or other defects—is crucial for ensuring product quality. However, the limited availability of annotated images of such defects poses challenges for developing reliable detection models. This paper proposes a novel image augmentation algorithm to address this limitation by generating synthetic imperfections from a small sample set. Unlike traditional approaches that augment entire images, the proposed approach isolates defects and applies spatial transformations (e.g., rotation, shear) before blending them into background images. This allows for the creation of diverse training datasets. The algorithm employs a 2D binary map to encode the spatial relationship between defect and background regions of images and then applies an AI-driven model to generate new realistic blended images. By leveraging a generator-discriminator framework, the augmentation process is iteratively refined during training to maintain the style and texture of the original imperfections. Experimental results show that the proposed method significantly enhances the performance of defect detection models compared to existing augmentation techniques while preserving the realism of the generated images.

Název v anglickém jazyce

Anomaly-Focused Augmentation Method for Industrial Visual Inspection

Popis výsledku anglicky

In industrial inspection, the detection of surface defects—such as scratches, dents, or other defects—is crucial for ensuring product quality. However, the limited availability of annotated images of such defects poses challenges for developing reliable detection models. This paper proposes a novel image augmentation algorithm to address this limitation by generating synthetic imperfections from a small sample set. Unlike traditional approaches that augment entire images, the proposed approach isolates defects and applies spatial transformations (e.g., rotation, shear) before blending them into background images. This allows for the creation of diverse training datasets. The algorithm employs a 2D binary map to encode the spatial relationship between defect and background regions of images and then applies an AI-driven model to generate new realistic blended images. By leveraging a generator-discriminator framework, the augmentation process is iteratively refined during training to maintain the style and texture of the original imperfections. Experimental results show that the proposed method significantly enhances the performance of defect detection models compared to existing augmentation techniques while preserving the realism of the generated images.

Druh

J_imp - Článek v periodiku v databázi Web of Science

CEP obor

—

OECD FORD obor

20200 - Electrical engineering, Electronic engineering, Information engineering

Projekt

<a href="/cs/project/EF16_019%2F0000867" target="_blank" >EF16_019/0000867: Centrum výzkumu pokročilých mechatronických systémů</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název periodika

IEEE Access

ISSN

2169-3536

e-ISSN

2169-3536

Svazek periodika

2025

Číslo periodika v rámci svazku

Not specified

Stát vydavatele periodika

US - Spojené státy americké

Počet stran výsledku

25

Strana od-do

10-15

Kód UT WoS článku

001498285200030

EID výsledku v databázi Scopus

—