Few-Shot Data Augmentation for Enhanced Surface Defect Detection in Industrial Manufacturing
Paper Title: Few-Shot Data Augmentation for Enhanced Surface Defect Detection in Industrial Manufacturing
Journal Name: Journal of Communications and Networks(JCN)
Abstract: Detecting defects on steel surfaces is essential for
maintaining production quality and reducing material waste.
However, the limited availability of defect images presents a
significant challenge in building accurate and reliable detection
models. To overcome this limitation, we introduce DefectGen+,
an innovative data augmentation technique that utilizes Stable
Diffusion and blending methods to enhance defect image datasets,
thereby improving detection accuracy. Our approach aims to
generate a larger, more realistic dataset from a small number
of defect images. The image generation process consists of two
stages: the Defect-Free Stage and the Defect Generation Stage.
In the first stage, Stable Diffusion is trained on a dataset free
from defects to create realistic images of steel surfaces. In the
Defect Generation Stage, which is further divided into two
phases, defects are first extracted and modified from a limited
set of defect images. In the second phase, these modified defects
are seamlessly blended into the defect-free images at specified
locations, producing highly realistic defect images. Our experience
ments on a steel surface dataset show that DefectGen+ enhances
detection performance, achieving an 11.91% improvement in
Mean Average Precision (mAP) for defect detection.