【论文笔记】CVPR三篇HIC论文

随便找个文章记录一下：classification reports里的recall才是CA

1 Hyperspectral image classification based on multi-scale residual network with attention mechanism

Zhang, Xiangdong, Tengjun Wang and Yun Yang. “Hyperspectral Images Classification Based on Multi-scale Residual Network.” (2020).

1.1 具体方法

  为了提取有效信息，忽略无效信息，本文提出了一种引入attention机制的多尺度残差网络。

1. 整体网络架构
   
   
   （跟HybridSN类似啊，先是3D CNN，然后是2D CNN）
2. Spectral FE Block、Spatial FE Block
   
3. SSA（Spatial-Spectral attention）
   
4. SA（Spectral attention）
   

1.2 实验

  有个值得学习的实验——将Attention设在Conv block前面还是后面呢？论文比较了Attention-PRE、Attention-EMBED、Attention-POST三种策略的分类精度。

  注意到不同数据集下三种策略的分类精度都差不多。

2 CONVOLUTION BASED SPECTRAL PARTITIONING ARCHITECTURE FOR HYPERSPECTRAL IMAGE CLASSIFICATION

Chu, Ringo S. W., Ho-Cheung Ng, Xiwei Wang and Wayne Luk. “Convolution Based Spectral Partitioning Architecture for Hyperspectral Image Classification.” IGARSS 2019 - 2019 IEEE International Geoscience and Remote Sensing Symposium (2019): 3962-3965.
我不知道为啥在arxiv上显示CVPR，结果在IEEE上是IGARSS。

2.1 具体方法

Operations of each segments are identical, kernel weights and biases of the stack of convolutions are shared across the segments to reduce total number of parameters.
（每个segments（其实也就两个）后面都是4个3D conv操作，它们的权值和偏置都是共享的，这样可以减少参数）

3 Multitask Deep Learning with Spectral Knowledge for Hyperspectral Image Classification

S. Liu and Q. Shi, “Multitask Deep Learning With Spectral Knowledge for Hyperspectral Image Classification,” in IEEE Geoscience and Remote Sensing Letters, doi: 10.1109/LGRS.2019.2962768. 我不知道为啥在arxiv上显示CVPR，结果在IEEE上是GRSL。

3.1 具体方法

With the spectral knowledge, the first convolutional layer is weight sharing; without the spectral knowledge, each data set has its own weights of the first convolutional layer.（有光谱知识，指不同数据集之间的第一层卷积层共享参数；无光谱知识则不共享）

3.2 实验

  在SV上的表现不尽如人意呀。

4 单词（Je vous connais depuis toujours.）

今日推荐——欢乐好声音 Sing (2016)
2020-6-12 09:27:53
你知道跌入谷底的好处是什么吗？
you know the good thing about hitting rock bottom？