基于生成对抗网络的气缸套表面缺陷图像生成

针对气缸套缺陷检测中缺陷样本不足限制气缸套缺陷检测性能提升问题,采用基于生成对抗网络的气缸套表面缺陷检测算法.首先,为了保持缺陷图像中原有缺陷位置与特征不变,通过循环生成对抗网络模型学习有缺陷气缸套图像与正常图像的关系;其次,利用学习得到的模型对有缺陷气缸套图像进行风格迁移,即把有缺陷气缸套图像背景替换成无缺陷气缸套图像背景,实现对气缸套缺陷数据集的扩充与增强;最后,通过基于数据增强的RetinaNet网络模型对生成图像的有效性进行验证.试验结果表明,通过生成对抗网络生成的气缸套数据集可以提升缺陷检测性能,进一步证明了生成对抗网络在工业应用的可行性.

Generation of Cylinder Liner Surface Defect Images Based on Generative Adversarial Network

To address the lack of defect samples limits the improvement of detection performance in cylinder liner defect detection, this article adopts a cylinder liner surface defect detection algorithm based on Generative Adversarial Networks. Firstly, in order to maintain the original defect positions and features in the defect images unchanged, this paper uses a cyclic Generative Adversarial Network model to learn the relationship between defect images and normal images; Secondly, using the learned model to transfer the style of images with defects, that is, replacing the background of images with defect free images, to expand and enhance the cylinder liner defect dataset. Finally, this paper uses a RetinaNet network model based on data augmentation for surface defect detection of cylinder liners. The experimental results show that the cylinder liner dataset generated by Generative Adversarial Networks can effectively improve defect detection performance, further proving the feasibility of Generative Adversarial Networks in industrial applications.