导航

ACTA AERONAUTICAET ASTRONAUTICA SINICA ›› 2022, Vol. 43 ›› Issue (3): 425187.doi: 10.7527/S1000-6893.2021.25187

Previous Articles     Next Articles

Detection method of defects in automatic fiber placement based on fusion of infrared and visible images

KANG Shuo1, KE Zhenzheng2, WANG Xuan1, ZHU Weidong1   

  1. 1. College of Mechanical Engineering, Zhejiang University, Hangzhou 310027, China;
    2. Institute of Advanced Technology, Zhejiang University, Hangzhou 310027, China
  • Received:2020-12-31 Revised:2021-01-11 Online:2022-03-15 Published:2021-02-18
  • Supported by:
    Pioneer and "Leading Goose" Research and Development Program of Zhejiang (2022C01134)

Abstract: To improve the quality of Automatic Fiber Placement (AFP), this paper analyzes the limitations of single spectrum detection technology in the field of visual inspection, and proposes a method based on deep learning for defect detection by fusion of infrared and visible images to realize detection, location and classification of fiber placement defects.According to the difference between the defects in thermal infrared image and visible image, the feature fusion network is used to fuse the two kinds of spectral information to improve the detection effect.To improve the detection speed, the single-stage detection network is used as the detection framework.We analyze the shortcomings of anchor-based method in the single-stage detection network according to the characteristics of seriously uneven distribution of length-width ratio of fiber defects, and propose a detection method with anchor-free network, in which an improved feature pyramid network structure is added for multi-scale prediction.Using mean Average Precision (mAP) as the measurement index, the experimental results improved by 6.30%, 6.64% and 1.02% respectively compared with single spectrum detection method, anchor-based detection method and the detection method without improved feature pyramid structure.The detection time of each 608 pixels×608 pixels image is less than 20 ms after acceleration by Tensor RT, which meets the demand of real-time detection.The average recall of detection is more than 88%, and the average precision is more than 82%, which meets the demand of production accuracy.The data of this paper is verified by off-line detection on the test-bed, and is tested online on large gantry AFP equipment.

Key words: automatic fiber placement, defects detection, deep learning, machine vision, multi-spectrum fusion

CLC Number: