研究目的
To propose a ship detection method for large-scale images that does not require sea-land segmentation as a pre-processing step and can detect ships directly from a complicated background including sea and land.
研究成果
The proposed method effectively removes the interference of similar ground objects on land without the need for pre-sea-land segmentation, with optimal detection accuracy achieved when the proportion of positive and negative sample images is 1:1.
研究不足
The method's performance may be affected by the proportion of negative sample images used in training, with too many negative samples leading to missed detections.
1：Experimental Design and Method Selection:
Based on Faster R-CNN, the method uses a region proposal network (RPN) and fast R-CNN (FRN) stages for ship detection without sea-land segmentation.
2：Sample Selection and Data Sources:
Large-scale images from GF-1 and GF-2 satellites are used, including positive sample images with ships and negative sample images with only land areas.
3：List of Experimental Equipment and Materials:
i5-7500 CPU, 16 GB RAM hardware, 6 GB Graphics card of NVIDIA 1060, and the TensorFlow framework under Linux.
4：Experimental Procedures and Operational Workflow:
Training the network with a mix of positive and negative sample images, testing on large-scale images, and evaluating detection accuracy.
5：Data Analysis Methods:
Calculation of recall and precision based on the number of correctly detected ships.
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
