Shearography-Based Near-Surface Defect Detection in Composite Materials: A Spatiotemporal Object Detection Neural Network Trained Only with Simulated Data

Shearography is a non-destructive defect detection technique that, when combined with neural networks, can efficiently and accurately detect near-surface defects in composite materials. However, the high cost of the dataset significantly limits the application of neural networks in shearography. Current simulation data generation techniques fail to eliminate the discrepancies between simulated and experimental data, resulting in suboptimal performance when training neural networks with only simulated data. To address this issue, this paper utilizes phase map sequences measured by shearography as the medium for defect detection and designs a YOWO_SS3D spatiotemporal object detection network. The network simultaneously learns both the spatial distribution features and temporal variation patterns of simulated phase map sequences, achieving high-accuracy detection of defects. The experimental results show that, with only 4000 frames of simulated data for training, our network achieved a detection accuracy of 96.99% on experimental phase maps, which is considerably higher than the 65.37% accuracy achieved by training the YOLOv4 network with the same simulated data. Using our technique, only pre-generated simulation data are required to train the network, enabling YOWO_SS3D to be directly deployed for practical defect detection tasks. This approach eliminates the substantial costs associated with collecting experimental training data and promotes the application of neural network technology in the shearography field.