Unsupervised Learning-Based Method for Surface Defect Detection of Carbon Fiber Prepregs

Abstract Carbon fiber prepreg is the key material for preparing carbon fiber composites, and its surface defects directly affect the performance of composites. Existing supervised learning defect detection methods require predefined defect types and rely on a large number of defect labeled samples, making it difficult to adapt to the challenges of unknown defects due to changes in the production environment. An unsupervised anomaly detection method for surface defects of carbon fiber prepregs is developed and the corresponding carbon fiber prepreg surface anomaly detection model (CFP-AD) is constructed in this paper, which only needs to use merely non-defective images for training. Firstly, the various features taken from VGG19 are fed into the designed multi-branch dimensional cross-refinement feature fusion module, which enhances the representation ability of multi-scale convolutional feature. Secondly, a convolutional autoencoder with residual pooling is designed to efficiently realize feature reconstruction, improving the reconstruction ability of the CFP-AD model for the edges of carbon fiber prepregs. Lastly, a context semantic similarity measurement function is introduced based on the l2 norm anomaly score measurement to make the CFP-AD model more effective in highlighting defective regions. Extensive experimental results indicate that the CFP-AD model is able to accurately identify the image anomaly and locate defective regions in carbon fiber prepregs.