研究目的
To solve the inshore ship detection problem by combining global and local information of SAR images using a multi-scale full convolutional network (MS-FCN) and a rotatable bounding box based object detection method (DR-Box).
研究成果
The proposed method performs well on the task of inshore ship detection in SAR images by applying multi-scale information in segmentation and angle information in detection. It successfully combines the ability of accurately locating local pixels on a ship while remaining perception of global information.
研究不足
The method's effectiveness is demonstrated on specific SAR images (Chinese Gaofen-3 satellite images), and its generalizability to other SAR images or conditions is not explored. The computational complexity and memory requirements for processing large SAR images (often larger than 10000×10000) are not addressed.
1：Experimental Design and Method Selection:
The study employs a pipeline procedure combining a semantic segmentation network (MS-FCN) and an object detection network (DR-Box) for inshore ship detection in SAR images.
2：Sample Selection and Data Sources:
Training and testing are conducted using Chinese Gaofen-3 satellite images, including two imaging modes with different resolutions (8m and 3m).
3：List of Experimental Equipment and Materials:
The study utilizes deep neural networks (DNNs) for feature representation and extraction.
4：Experimental Procedures and Operational Workflow:
The input image is divided into a pyramid structure, processed through FCNs at different scales, and then merged to achieve sea-land segmentation. DR-Box is then applied on the sea region for ship detection.
5：Data Analysis Methods:
The performance of the proposed method is evaluated based on its ability to accurately locate inshore 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
