기계학습을 이용한 색상형 미세조직의 결정립 크기 측정 (Machine Learning-Based Prediction of Grain Size from Colored Microstructure)

서지정보

· 발행기관 : 대한금속·재료학회
· 수록지 정보 : 대한금속·재료학회지 / 61권 / 5호 / 379 ~ 387페이지
· 저자명 : 정준호, 김희수

초록

We constructed a convolutional neural network to estimate average grain size from microstructureimages. In the previous study from our research group, the network was trained using GB-type images inwhich the grain matrix and grain boundary were represented in white and black, respectively. The model wellestimated the same GB-type images, but did not properly predict CL-type images where grain boundarieswere defined by color contrast between grains. In the present study, the convolutional neural network wastrained using CL-type microstructure images, and evaluated the average grain size, for comparison with theprevious GB-type model. The relationship between the microstructure image and the average grain size wasdetermined using regression rather than classification. Then, the results were compared with the previousones. Finally, the proposed approach was used for actual microstructural image analysis. Mid-layer imageswere extracted to examine how the network recognizes the characteristics of microstructures, such as graincolor and grain boundary. Like the previous GB-type model, the present CL-type model seems to estimatethe average grain size from the curvature of the grain boundary through edge detection of the grainboundaries. However, the GB- and CL-type models only properly predicted the grain size from the same kindof images as the training data, because the definitions of the grain boundaries of the two models weredifferent.

영어초록

We constructed a convolutional neural network to estimate average grain size from microstructureimages. In the previous study from our research group, the network was trained using GB-type images inwhich the grain matrix and grain boundary were represented in white and black, respectively. The model wellestimated the same GB-type images, but did not properly predict CL-type images where grain boundarieswere defined by color contrast between grains. In the present study, the convolutional neural network wastrained using CL-type microstructure images, and evaluated the average grain size, for comparison with theprevious GB-type model. The relationship between the microstructure image and the average grain size wasdetermined using regression rather than classification. Then, the results were compared with the previousones. Finally, the proposed approach was used for actual microstructural image analysis. Mid-layer imageswere extracted to examine how the network recognizes the characteristics of microstructures, such as graincolor and grain boundary. Like the previous GB-type model, the present CL-type model seems to estimatethe average grain size from the curvature of the grain boundary through edge detection of the grainboundaries. However, the GB- and CL-type models only properly predicted the grain size from the same kindof images as the training data, because the definitions of the grain boundaries of the two models weredifferent.

참고자료

· 없음