CNN卷积算法的改进

改进有：空洞卷积、可变形卷积

（1）空洞卷积：对于像素要求不严格的任务，感受野相当于普通3*3卷积的两层的效果。

代码实现：

def DilatedCNN(x):

        length=len(x,filter)

        sum=0

        if length<5:

               return 0

        if (length-3)%2==0:

               for I in range(length+1):

                      x[length][i]=0

                      x[i][length]=0

        next_length=(length-3)/2

        for i in range(next_length):

               for j in range(next_length):

                      for k in range(3):

                             for n in range(3):

                                    sum+=x[i+2k][j+2n]*filter[k][n]

                      y[i][j]=sum

                      sum=0

        return y

（2）可变形卷积

          1）对于像素分布不是等距或者是立体像素分布的情况下，采用最邻近的n个点卷积的计算方式；

          2）对于平面均匀分布的像素点，卷积点在原先的基础上做了微调

还有一种卷积区域随机变化的可变形卷积：

代码实现如下：

def DeformableCNN(x):

        length=len(x,filter)

        sum=0

        if length<5:

               return 0

        if (length-3)%2==0:

               for I in range(length+1):

                      x[length][i]=0

                      x[i][length]=0

        next_length=(length-3)/2

        for i in range(next_length):

               for j in range(next_length):

                      for k in range(3):

                             for n in range(3):

                                    row=shuffle[0,1,2,3,4] //0-4的数随机排序

                                    col= shuffle[0,1,2,3,4]  //0-4的数随机排序

                                    sum+=x[i+row[k]][j+col[n]]*filter[k][n]

                      y[i][j]=sum

                      sum=0

        return y

（3）反卷积：即卷积的逆过程。

我设计的DeformableCNN发表了一篇论文，还未出版：A  Modulation Classification Method Based on Deformable Convolutional Neural Networks for Broadband Satellite Communication Systems