Robust defect image synthesis using null embedding optimization for industrial applications

초록

Accurate defect classification and segmentation are critical in the manufacturing sector, yet both tasks are often hindered by imbalanced data and the scarcity of defect samples. Traditional synthetic data augmentation methods tend to produce images with structural inconsistencies, limiting their effectiveness. In this work, we introduce a novel approach that integrates null embedding optimization with Residual Linear Interpolation (RLI) connections to generate latent representations that closely mimic the original images while preserving structural fidelity. Furthermore, a prompt-to-prompt augmentation technique is employed to systematically modify the base text prompt, enabling the generation of diverse defect morphologies. This unified framework primarily enhances the variability of the dataset by generating diverse defect morphologies, while simultaneously yielding high-fidelity synthetic images that visually correspond to real defects, thereby significantly improving the performance of both classification and segmentation models. The source code and models are available at https://acerghjk-cloud. github.io/ESWA2025/

키워드