Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (18): 234-241.DOI: 10.3778/j.issn.1002-8331.2205-0583

• Graphics and Image Processing • Previous Articles     Next Articles

Image Style Transfer Method for Industrial Inspection

ZHU Zhongxian, MAO Yushi, CAI Kewei, LIU Wentao, PU Daojie, DU Yao, WANG Zilei   

  1. 1.Ultra High Voltage Company, State Grid Anhui Electric Power Co., Ltd., Hefei 230061, China
    2.Institute of Advanced Technology, University of Science and Technology of China, Hefei 230031, China
  • Online:2023-09-15 Published:2023-09-15



  1. 1.国网安徽省电力有限公司 超高压分公司,合肥 230061
    2.中国科学技术大学 先进技术研究院,合肥 230031

Abstract: For the inspection tasks in industrial scenarios such as electric power, it is necessary to convert the virtual images generated by virtual simulation into real style to assist the next photovoltaic panel positioning and defect detection. However, most of the traditional image style transfer methods focus on the translation between artistic image and real scene. Due to the characteristics of artistic image itself, to a certain extent, the content structure of image may change before and after translation. This change will have a negative impact on the completion of downstream tasks, so it can hardly be applied in industrial scenarios. A contrastive learning-based visual style transfer method is proposed in this research. The created images can be closer to the target domain in the deep features by comparing the features of the same and different image blocks on many layers of features, while restricting the image before and after the transfer. The image content is uniform. The experimental results suggest that the strategy proposed in this study can reduce model training time and computational overhead while assuring that the key content of image is not warped or absent.

Key words: image style transfer, generative adversarial network, virtual reality, contrastive learning

摘要: 对于电力等工业场景中的巡检任务,需要将虚拟仿真生成的虚拟图像转换为真实的风格,辅助接下来的设备定位和设备缺陷检测工作。然而传统的图像风格迁移方法大多着眼于艺术图像与现实场景的转换,由于艺术图像本身的特性,在迁移前后图像的内容结构可能会出现一定程度的变化,这种变化会对下游任务的性能造成负面影响,因而无法直接应用。因此提出一种基于对比学习的图像风格迁移方法,通过在多个层次的特征上分别对比相同与不同位置图像块的特征,能够使生成图像在深层特征上向目标域靠近,同时约束迁移前后的图像内容上保持一致。实验表明,所提出的方法能够在保证图像主体内容不发生变形和缺失的同时,节省模型训练过程中的计算开销。

关键词: 图像风格迁移, 生成对抗网络, 虚拟仿真, 对比学习