Predicting Gradient is Better Exploring Self-Supervised Learning for SAR ATR with a Joint-Embedding Predictive Architecture

• 论文名称: Predicting Gradient is Better: Exploring Self-Supervised Learning for SAR ATR with a Joint-Embedding Predictive Architecture
• Link: https://arxiv.org/abs/2311.15153v4
• Published in: Arxiv 2023
• Type: Pretrain
• Code/Project: https://github.com/waterdisappear/SAR-JEPA
• 备注: 自监督，类似JEPA的框架，固定student生成梯度特征
• 数据类型: SAR
• Backbone: ViT
• 下游任务: Target Recognition
• Short Summary: 发现像素采样不适用于SAR图像，因为SAR图像的单个像素包含乘性噪声。因此更倾向于使用局部补丁进行遮挡，而不是整个图像或像素级别。工作旨在通过局部补丁在目标层次上挖掘语义信息，而不是场景层次，工作发现梯度比率比HOG的差分梯度在乘性斑点噪声下更适合目标形状信息提取。