Aircraft target recognition in remotely sensed imagery has important applications in areas such as military reconnaissance, airport management, aviation safety and disaster emergency response. The main challenges lie in dealing with complex backgrounds, multi-scale variations and high-resolution data. To improve the recognition accuracy and efficiency, we propose a lightweight keypoint detection model based on YOLOX improvement. First, we replace the backbone structure in the YOLOX network with the ShuffleV2 structure, which ensures efficient information delivery through channel separation and merging, while reducing the memory access cost and accelerating inference on edge computing devices with limited computing power. Second, we adopt PAFPN (Path Aggregation Feature Pyramid Network) spatial feature pyramid structure to fuse the input feature maps of the network backbone to minimize the computational load and memory overhead. Finally, we modify the decoupling header of YOLOX to predict the coordinates of the four vertices of the bounding box as well as the five critical points of the aircraft target. This modification significantly enhances the accuracy and robustness of the recognition system. Experimental results show that the overall average accuracy of the modified YOLOX model is $\mathbf{9 9. 4 \%}$, which is $\mathbf{2 1 \%}$ better than the original YOLOX model. In addition, the improved model has $57 \%$ fewer parameters, $48.7 \%$ fewer Flops, $\mathbf{2 5. 6 \%}$ less inference time, and $\mathbf{9. 6 \%}$ less memory usage. This lightweight model significantly reduces computational resource requirements while maintaining high detection accuracy, improving model speed and applicability for resource-constrained environments such as mobile devices and embedded systems.
Research on Aircraft Target Recognition in Remote Sensing Images Based on YOLOX Model
23.10.2024
1583484 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
YOLOX-based dynamic vehicle target detection
IEEE | 2022
|Research on YOLOX-based tire defect detection method
IEEE | 2022
|