A Generative Adversarial Network-Based Method for Intelligent Detection of Military Targets with Few Training Samples for Embedded Edge Devices

Abstract

With the development of artificial intelligence and intelligent algorithms, new technical methods of automatic military target recognition in complex environments have been provided. However, limited availability of labeled samples, significant scale variations in targets and computing resource limitations on embedded devices all still limit detection accuracy and the ability to deploy detection in real-time. In this regard, a generative adversarial network (GAN) based few-shot military target detection is presented. To enlarge the training data set, class-conditional sample generation, multi-scale feature encoding, target-region attention constraints and generated-sample quality screening are employed, and the joint training of real and generated samples is used to enhance the generalization ability of the model. Based on this, depthwise separable convolution, channel pruning, structural re-parameterization, mixed-precision quantization, knowledge distillation, and operator fusion are introduced to realize lightweight network design and optimize the network on the edge. Experimental results show that the proposed method achieves a precision of 91.2%, a recall of 88.9%, and an F1-score of 90.0%. The mAP@0.5 and mAP@0.5:0.95 reach 92.7% and 61.8%, respectively, while the model contains only 4.1 M parameters and requires 7.6 G FLOPs. On the NVIDIA Jetson Orin NX platform, the inference speed reaches 106.4 FPS. These results demonstrate that the method effectively balances few-shot detection accuracy, robustness in complex environments, and real-time inference requirements on embedded edge devices.