针对目前深度学习检测算法在遥感图像在轨实时处理任务上综合性能不平衡，且在星载嵌入式设备上难以部署的问题，本文基于YOLOX-s算法，提出一种面向嵌入式平台的光学遥感飞机目标快速检测算法LAD-YOLOX（Lightweight Aircraft Detection YOLOX）。首先，在硬件感知网络的设计上，基于ShuffleNetv2设计了超轻量、高精度的骨干模块ES-Block（Enhanced ShuffleNet Block），重构原有的主干特征提取网络。其次，引入GSConv构建轻量级颈部特征融合网络GS-Neck，使得前后端结构的参数量配比均衡，降低了计算复杂度的同时精度损失更小。然后，设计轻量级解耦合检测头网络结构，进一步提升飞机目标分类和定位的特征编码，降低模型参数量，提升检测性能。最后，为LAD-YOLOX算法替换置信度预测损失函数Varifocal Loss及边界框定位损失函数SIoU Loss，提高了模型训练的收敛速度和推理时的精度。仿真实验结果表明，所提算法在RSOD遥感数据集上，检测精度损失极低，计算量压缩为原始YOLOX-s模型的43.72%，检测速度提高24 FPS。在XILINX EK-U1-ZCU102-G评估套件上完成算法部署及加速，对自制飞机数据集的检测速度至少可达到26.53 FPS，能够满足在轨实时、准确检测飞机目标的需求。

Objective With the rapid development of space remote sensing technology, the number of high-resolution optical remote sensing images is increasing exponentially. The detection of high strategic value targets such as aircraft is currently a hot research topic in the field of high-resolution image processing for remote sensing. Traditional remote sensing image object detection algorithms include template matching and traditional machine learning. These algorithms mostly rely on prior knowledge from experts, and the features used are mostly primary manual features limited to pixel level, which have certain limitations and cannot cope with complex and ever-changing backgrounds as well as multimodal and diverse targets. As regards to deep learning technology, remote sensing object detection algorithms include two-stage methods and one-stage methods. Two-stage method has high accuracy, but it consumes a lot of resources and limits processing speed. The YOLO detection algorithms are widely concerned and applied due to their simple network structure, balancing detection accuracy and speed. However, it is difficult for one-stage models to be directly deployed on embedded devices on satellites for real-time detection of aircraft targets, due to the limitations of computing power, storage capacity, and model complexity. Therefore, there is an urgent need to study the lightweight network models which can reduce their demand for computing power and storage. The network models with excellent target detection capabilities are then deployed to aerospace chips with limited resources to complete efficient aircraft target detection tasks. Method To solve the problem that it is difficult to deploy current network models with excellent target detection capabilities to aerospace chips with limited computing and storage resources, this paper proposes five targeted designs for the benchmark model which is based on the one-stage YOLOX-s algorithm, and the implementation of the model adopts lightweight design concept. A fast optical remote sensing aircraft target detection algorithm LAD-YOLOX (Lightweight Aircraft Detection YOLOX) has been proposed which is suitable for deployment on embedded platforms. By testing the RSOD dataset and self-made aircraft dataset, the effectiveness of the improved strategy and the generalization of the overall model were verified, achieving extremely low-loss compression of the model. Finally, algorithm deployment and acceleration were completed on the XILINX EK-U1-ZCU102-G evaluation kit, which can meet the requirements of real-time performance and detection accuracy. Result First, during the ablation experiment stage, the effectiveness of each improvement module is verified by gradually accumulating the five improvement strategies which are based on the baseline YOLOX-s. Compared to the YOLOX-s, LAD-YOLOX achieves extremely low-loss model compression. The parameter quantity is reduced to 58.83% of the benchmark, and the calculation quantity is reduced to 43.72% of the benchmark, mAP@.5(%) and mAP@.5:.95(%) only reduced by 0.2%, and the detection speed increased by 24 FPS. Then, the proposed design is applied to the detection task of a self-made aircraft dataset. When the depth and width configuration is set to 0.33 and 0.375, the computational cost is 35.45 GFLOPs, and 54.04% mAP is obtained. FP16 bs=1 is used for inference on RTX4090, the speed can reach 59.50 FPS, which is the best comprehensive performance model in the comparative experiment. Finally, using Quantitative Aware Training, the model was deployed and tested on the ZCU102 evaluation kit to obtain the visualization results and detection speed of aircraft targets. When using 6144×6144 images as input, the average processing speed can reach 26.53 FPS, and the aircraft targets annotated in the self-made aircraft dataset can be accurately identified. Some aircraft targets have overlapping detection frames, misjudgment of difficult cases, and false negatives. Conclusion In response to the issue of imbalanced comprehensive performance of deep learning detection algorithms in real-time processing of remote sensing images in orbit, and difficulty in deployment on embedded devices on satellites, according to the YOLOX-s algorithm, this paper proposes a fast aircraft detection algorithm named LAD-YOLOX (Lightweight Aircraft Detection YOLOX) which based on optical remote sensing for embedded platforms. First, an ultra-lightweight and high-precision backbone module ES-Block is designed on hardware-aware network to reconstruct the original backbone, which based on ShuffleNetv2. Second, GSConv is introduced to construct lightweight neck which is named GS-Neck. This design equalizes the parameters between the front and the back, reducing the computational complexity and the precision loss. Then, the lightweight decoupling detection head is designed to further improve the feature coding of classification and location, reducing the parameters to improve the detection performance. Finally, Varifocal Loss and SIoU Loss are used in the loss function of LAD-YOLOX, which improves the convergence speed of model training and the precision of reasoning. The simulation results show that the proposed algorithm on the RSOD dataset makes the detection accuracy lossless and the calculation quantity is compressed to 43.72% of the original YOLOX-s model. The detection speed is increased by 24 FPS. The algorithm deployment and acceleration are completed on XILINX EK-U1-ZCU102-G evaluation kit. The detection speed of the self-made aircraft dataset can reach 26.53 FPS, which meets the requirements of real-time and accurate detection of aircraft targets in orbit.