고해상도 콘택트렌즈 이미지의 크랙 검출을 위한Asymmetric Residual U-Net
Asymmetric Residual U-Net for Crack Detection in High-Resolution Contact Lens Images
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
본 연구에서는 고해상도 콘택트렌즈 이미지에서 미세 크랙을 검출하기 위해 Asymmetric Residual U-Net 모델을 제안한다. 이모델은 인코더에서 Asymmetric Residual Convolutions를 활용하여 작은 객체의 미세 특징을 정밀하게 학습하고, 디코더를 최적화하여 연산량을 줄이는 동시에 Attention 모듈을 통해 정보 손실을 완화하였다. 또한, 다중 스케일 정보를 통합하기 위해 SpatialPyramid Pooling(SPP)를 적용하여 크랙 검출 성능을 향상시켰다. Focal Tversky Loss와 Edge-aware Weights를 결합하여 클래스 불균형 문제를 효과적으로 해결하였으며, 실험 결과, 본 모델은 기존 U-Net 대비 4.18%의 성능 향상을 보였다. 본 연구는 렌즈 제조 공정에서 미세 크랙 검출의 정확성과 효율성을 개선할 수 있음을 입증하였다.
Abstract
In this research, we propose an Asymmetric Residual U-Net model for detecting micro-cracks in high-resolutioncontact lens images. The model utilizes Asymmetric Residual Convolutions in the encoder to effectively capturefine-grained features of small objects, while optimizing the decoder to reduce computational complexity andmitigate information loss through Attention modules. Additionally, an Spatial Pyramid Pooling (SPP) moduleintegrates multi-scale information to enhance crack detection performance. By incorporating Focal Tversky Losswith Edge-aware Weights, the model addresses class imbalance more effectively. Experimental results demonstratea 4.18% performance improvement over the baseline U-Net. This work marks a significant advancement in theaccuracy and efficiency of micro-crack detection within lens manufacturing processes.
참고문헌 (8)
- Ronneberger, Olaf, "U-net : Convolutional networks for biomedical image segmentation", Medical image computing and computer-assisted intervention, pp.234~241 (2015)
- Zhou, Zongwei, "Unet++ : A nested u-net architecture for medical image segmentation", Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support:4th International Workshop, pp.3~11 (2018)
- Z. Zhang, "Road Extraction by Deep Residual U-Net", IEEE Geoscience and Remote Sensing Letters, 15(5), pp.749~753 (2018)[DOI]
- Jha, Debesh, "Resunet++ : An advanced architecture for medical image segmentation", 2019 IEEE international symposium on multimedia, pp.225~2255 (2019)
- Hu, Jie, "Squeezeand-excitation networks", Proceedings of the IEEE conference on computer vision and pattern recognition, pp.7132~7141 (2018)
- Vaswani, A, "Attention is all you nee", Advances in Neural Information Processing Systems (2017)
- He, Kaiming, "Spatial pyramid pooling in deep convolutional networks for visual recognition", IEEE transactions on pattern analysis and machine intelligence, 37(9), pp.1904~1916 (2015)[DOI]