전이학습 및 GAN 기반 데이터 증강을 통한 미세형상 사출부품의 향상된 결함 분류
Enhanced Defect Classification for Micro-Shaped Injection Molded Parts via Transfer Learning and GAN-Based Data Augmentation
3
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study presents an improved AI-based inspection system for micro-shaped injection molded parts used in smartphone actuators. Building on previous studies utilizing CNN-based transfer learning, this study focuses on maximizing the test accuracy by applying optimized imaging conditions, background variations, and advanced deep learning architectures. By designing a previously built lightweight PET jig and a semi-automatic image acquisition module that collects 11.94-megapixel RGB images, switching from 331x331 input resolution to 512x512 resolution and replacing the green background with a white metal jig, the baseline test accuracy is improved from 94.27% to 97.38%. The accuracy is further improved to 98.97% by using EfficientNet-B7 and Swin Transformer. In addition, the accuracy is improved to 99.42% by expanding the dataset through GAN-based image augmentation using StyleGAN2-ADA. These results demonstrate approximately 5% performance improvement over existing models, demonstrating the potential for high-precision, real-time defect classification suitable for mass production environments.
Abstract
This study presents an improved AI-based inspection system for micro-shaped injection molded parts used in smartphone actuators. Building on previous studies utilizing CNN-based transfer learning, this study focuses on maximizing the test accuracy by applying optimized imaging conditions, background variations, and advanced deep learning architectures. By designing a previously built lightweight PET jig and a semi-automatic image acquisition module that collects 11.94-megapixel RGB images, switching from 331x331 input resolution to 512x512 resolution and replacing the green background with a white metal jig, the baseline test accuracy is improved from 94.27% to 97.38%. The accuracy is further improved to 98.97% by using EfficientNet-B7 and Swin Transformer. In addition, the accuracy is improved to 99.42% by expanding the dataset through GAN-based image augmentation using StyleGAN2-ADA. These results demonstrate approximately 5% performance improvement over existing models, demonstrating the potential for high-precision, real-time defect classification suitable for mass production environments.