研究目的
To improve the surface accuracy of the FAST reflector by proposing a photogrammetry scheme for measuring the positions of 2226 nodes, with a focus on enhancing the node detection method in photos using a convolutional neural network (CNN) to achieve a higher recognition rate compared to traditional edge detection.
研究成果
The CNN-based method significantly improves node recognition rates to 91.5%, outperforming traditional edge detection (51.5%) and demonstrating robustness under varying lighting conditions. This advancement supports the feasibility of non-target photogrammetry for FAST reflector measurements, providing a foundation for accurate surface control. Future work will focus on DPU and camera calibration for precise node positioning in the object coordinate system.
研究不足
The method requires fixed regions for training to eliminate interference from image differences, which may limit applicability to other areas. The training process is complex and time-consuming, and the approach assumes known initial node positions for candidate region selection. Weather conditions at the FAST site (often cloudy and rainy) can affect image quality, though the CNN method shows robustness compared to traditional methods.
1：Experimental Design and Method Selection:
The study uses a convolutional neural network (CNN) with candidate regions for object detection, specifically adapting the LeNet5 structure for node recognition in grayscale images. The method involves training the CNN on a dataset of node and background images, with candidate regions selected based on initial node positions to reduce interference.
2：Sample Selection and Data Sources:
460 photos were taken of the FAST reflector around foundation pier 9 using a digital positioning unit (DPU) from foundation pier 6, capturing images under various lighting conditions (cloudy, sunny, rainy). The training set consists of approximately 22,000 node images and 48,000 background images extracted from these photos.
3：List of Experimental Equipment and Materials:
Equipment includes a digital positioning unit (DPU) with a camera, servo motors for rotation, and a control computer. The DPU is used for photogrammetry, with specifications including a lens with variable focal length and a high-precision rotating platform.
4：Experimental Procedures and Operational Workflow:
Photos were taken with the DPU, node coordinates were calculated using a central perspective projection model, and images were processed to extract node and background samples. The CNN was trained and validated, followed by node detection using sliding window and non-maximum suppression methods in candidate regions.
5：Data Analysis Methods:
The recognition rate was evaluated by comparing CNN-based detection with traditional edge detection under different lighting conditions and surface types (sphere and paraboloid). Statistical analysis of accuracy and loss functions was performed using tools like Tensorboard.
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
