Defect detection of laminated surface in the automated fiber placement process based on improved CenterNet

To solve the problem that the low defect detection accuracy on fiber laminated surface in Automated Fiber Placement(AFP) process based on visible images influenced by poor light source, low texture contrast of prepreg and other factors, a method for defect detection of laminated surface infrared images in AFP process based on improved CenterNet was proposed to improve defect detection performance on laminated surface in AFP process. First of all, due to the limit on hardware configuration of IPC, and large amounts of parameters on CenterNet model, lightweight MobileNetV3 network based on ASFF was utilized as the backbone network of AFP-CenterNet model to construct an anchor-free and lightweight detection model and reduce the number of network parameters and the occupancy of storage resources. Then, as for solving the bandwidth parameters of Gaussian kernel function, a method of adaptive adjustment of bandwidth parameters according to the aspect ratio of ground-truth bounding box was proposed to reduce the number of negative samples and loss error. Experimental results reveal that the improved AFP-CenterNet owns 90.2% AP in defect detection accuracy on the AFP infrared data set, 12.9 MB in model memory capacity, and 52 ms in detection time of a single sheet. Compared with the original backbone of CenterNet, detection accuracy of AFP-CenterNet is slightly worse than that of DLA-34, almost same with that of ResNet-101 and 7.7% higher than that of ResNet-18. Moreover, compared with DLA-34, ResNet-101 and ResNet-18, the model capacity of AFP-CenterNet decreased by 83.2%, 93.6% and 78.6% respectively. As for comparison with typical anchor-based network such as SSD and YOLOv3, AFP-CenterNet owns higher detection accuracy with specific 9.6% and 8.3%, and lower model capacity respectively reduced by 85.1% and 94.5%. Time spent on defect detection of AFP-CenterNet is nearly half that of CenterNet, SSD and YOLOv3 without using GPU to accelerate.