Defect detection of CFRP infrared thermal imaging based on multi-stream enhanced semi-supervised learning algorithm

A deep learning-based infrared nondestructive testing method is proposed for the internal defects or damages that may be formed in carbon fiber reinforced polymer during long-term service. The method collects infrared thermal radiation data of the specimen with the help of an optically-excited infrared thermal imaging system, combines with image preprocessing technology to construct a semi-supervised learning dataset, and employs a multi-stream-enhanced semi-supervised deep learning algorithm to solve the problems of data scarcity and labeling difficulty. In addition, the Pixel-Boosting framework is introduced to realize the accurate identification and segmentation of defective pixels through background pixel simulation, which significantly improves the detection performance in complex noise environments. Finally, the Pixel-Match algorithm is formed by combining the multi-stream enhancement algorithm and the Pixel-Boosting framework. The F-score value is used as the main evaluation index, and it is compared and analyzed with the fully supervised model (DeepLabV3+) and the semi-supervised model (U2PL). The experimental results show that the Pixel-Match algorithm achieves an average F-score value of 90.96%, which is 6.72% and 9.2% higher compared to DeepLabV3+ (84.24%) and U2PL (8ae162c1.76%), respectively, and demonstrates excellent robustness.