Herein, we propose a method to estimate fatigue damage of a thin metal film at the micro scale with high accuracy. With progress in the miniaturization of flexible printed circuits, estimation of damage at the micro scale is required to improve the reliability of industrial processing. Despite this necessity, it has been difficult to accurately evaluate the damage at the micro scale using conventional tests. Therefore, we focus on the use of microscopic images for damage estimation based on the correlation between damage and surface morphology, such as cracks. However, image-based analytical estimation by physical modeling is extremely difficult because of the complexity of the morphology. Therefore, the proposed method is based on image recognition using deep learning. In particular, VGG19-based transfer learning was implemented. An $L_2$-constrained softmax loss developed for face recognition was used, as the diversity of textures in the microscopic images was lower than that used for general object detection. As a result, the proposed method was able to accurately estimate the fatigue damage at the micro scale, which was at the submillimeter scale. The average estimation error was reduced to 15% of that obtained using the binarization method, indicating the $L_2$-constrained softmax loss method to be highly effective. Although verification under a broad range of fatigue conditions is required for a more general evaluation, it was concluded that the proposed method is effective for evaluating the damage of thin metal films of flexible printed circuits at the micro scale.

在此，我们提出了一种在微观尺度上高精度估计金属薄膜疲劳损伤的方法。随着柔性印刷电路小型化的进展，需要在微观尺度上估计损坏以提高工业加工的可靠性。尽管有这种必要性，但使用传统测试很难在微观尺度上准确评估损伤。因此，我们专注于基于损伤与表面形态（如裂纹）之间的相关性，使用显微图像进行损伤估计。然而，由于形态的复杂性，通过物理建模进行基于图像的分析估计是极其困难的。因此，所提出的方法基于使用深度学习的图像识别。特别是，实施了基于 VGG19 的迁移学习。一个${升}_2$使用为人脸识别开发的约束 softmax 损失，因为微观图像中纹理的多样性低于用于一般物体检测的纹理多样性。结果，所提出的方法能够在亚毫米尺度的微观尺度上准确估计疲劳损伤。平均估计误差降低到使用二值化方法获得的 15%，表明${升}_2$- 约束 softmax 损失方法非常有效。尽管需要在更广泛的疲劳条件下进行验证才能进行更一般的评估，但得出的结论是，所提出的方法对于在微观尺度上评估柔性印刷电路的金属薄膜损伤是有效的。