研究目的
To address the scalloping effect in ScanSAR images, which affects visualization and quantitative applications like surface wind and wave retrievals, by proposing a deep neural network based on residual learning for descalloping.
研究成果
The proposed DesNet based on deep residual learning effectively eliminates scalloping patterns in ScanSAR images, showing strong adaptive ability and superior performance compared to existing methods. Future work should focus on exploring better network architectures for the descalloping task.
研究不足
The method may have limitations in handling extremely complex image textures or scenarios not covered in the training dataset. Optimization of network architecture and loss function could be explored further.
1：Experimental Design and Method Selection:
A deep residual convolutional neural network (DesNet) with 56 layers is designed, using instance normalization and spatial reflection padding. The network is trained to map input ScanSAR images with scalloping to output descalloped images.
2：Sample Selection and Data Sources:
Dataset collected from GF-3 ScanSAR data, including narrow scan mode (NSC), wide scan mode (WSC), and global observation mode (GLO) images, with about 145k images of resolution 256*
3：List of Experimental Equipment and Materials:
2 Tesla K80 GPU for training, PyTorch framework for implementation.
4：Experimental Procedures and Operational Workflow:
Images are preprocessed by normalizing scalloping patterns to [-1,1]. Network is trained with a batch size of 32 for 64,000 iterations using Adam optimizer with specified learning rates and loss function factors.
5：1]. Network is trained with a batch size of 32 for 64,000 iterations using Adam optimizer with specified learning rates and loss function factors.
Data Analysis Methods:
5. Data Analysis Methods: Qualitative evaluation by visual inspection of output images, quantitative evaluation by calculating mean amplitude in selected subregions and plotting diagrams.
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
