[ConvLSTM] Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting阅读笔记

1、主要创新

              相比于FC-LSTM,引入卷积层,1)可以充分利用空间关联,2)能够输出图像序列,3)减少参数冗余

2、网络

        1)输入

                时空序列,特征数目=P,图像大小=MxN,即[ConvLSTM] Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting阅读笔记_第1张图片

                ​​​​​​​        ​​​​​​​        [ConvLSTM] Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting阅读笔记_第2张图片

        2)输出

                

        3)网络设计

                i: input gate, f: forget gate, c: cell status, h: final state, o: output gate,

                *: convolution operator, o: Hadamard product

        ​​​​​​​        ​​​​​​​        ​​​​​​​        ​​​​​​​[ConvLSTM] Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting阅读笔记_第3张图片

        ​​​​​​​        ​​​​​​​        ​​​​​[ConvLSTM] Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting阅读笔记_第4张图片

                -> 卷积核↑,捕捉速度↑

                -> FC-LSTM等价于h=w=1的ConvLSTM

                -> 需要0-padding,来保证state和input的大小一致

                -> 初始state设为0

你可能感兴趣的:(paper阅读笔记,深度学习,医学图像)