A Semisupervised Aircraft Fuselage Defect Detection Network With Dynamic Attention and Class-Aware Adaptive Pseudolabel Assignment

To track the problem of aircraft fuselage defect detection in complex environments and reduce aviation safety hazards such as careless observation and delayed reporting due to objective factors, a semi-supervised aircraft fuselage defect detection network was proposed. Firstly, we constructed a new baseline model that extends one-stage detector with dynamic head and partial convolution named as dynamic decoupled detector, which enhances the representation capability of the model and improves the detection accuracy of small defects. Secondly, to address the issue of inconsistent pseudo-label distribution in semi-supervised learning, we propose a class-aware adaptive pseudo-label assignment strategy that adaptively obtains the pseudo-label filtering threshold during the training iteration to further optimize the pseudo-label assignment process. Finally, to validate the effectiveness of the proposed model, we construct a data set for aircraft fuselage defect detection for semi-supervised training. Experimental results show that the proposed semi-supervised aircraft fuselage defect detection network outperforms the current state-of-the-art semi-supervised object detection framework on the aircraft fuselage defect dataset. At the same time, the proposed model has better generalization performance and provides more reliable support for real-time visualization of aircraft fuselage defects.