研究目的
To apply deep convolutional neural networks, specifically YOLOv2, for real-time and accurate detection of spacecraft in space environments, addressing limitations of existing methods in speed and accuracy.
研究成果
The modified YOLOv2 network achieves high detection accuracy (97.8% detection rate, 80.6% mAP) and fast processing times (0.018s per image), demonstrating effectiveness for real-time spacecraft detection with robustness to rotation and illumination changes. It outperforms YOLO and SSD, showing potential for wide applications in space target detection.
研究不足
The study relies on simulated and web-collected data, which may not fully represent real-space conditions. The dataset size is limited (2500 training images), and the method's performance in extremely complex or adversarial scenarios is not extensively tested.
1：Experimental Design and Method Selection:
The study uses a regression-based convolutional neural network (YOLOv2) for end-to-end spacecraft detection, with modifications to the architecture for improved performance. Data augmentation and fine-tuning techniques are employed to handle limited data.
2：Sample Selection and Data Sources:
A dataset of 2500 training images and 300 testing images is created using web-collected space target images and Blender software for simulation under different circumstances. Data augmentation includes rotation, inversion, and noise disturbance.
3：List of Experimental Equipment and Materials:
A personal computer with Intel Core i7 CPU, NVIDIA GTX-1080 GPU (8 GB memory), 16 GB RAM, running Ubuntu 16.04 OS. Software includes TensorFlow framework, Blender for image generation, and LabelImg for annotation.
4：04 OS. Software includes TensorFlow framework, Blender for image generation, and LabelImg for annotation.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Images are annotated and augmented, then input into the modified YOLOv2 network for training and testing. The network outputs bounding boxes and confidence scores for spacecraft detection.
5：Data Analysis Methods:
Performance is evaluated using mean average precision (mAP), detection rate, and average running time per image, with comparisons to YOLO and SSD methods.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
