如何在Python中从其补丁中恢复3D图像？

Question

我有一个形状为“DxHxW”的三维图像。我成功地提取了图像
分为“pdxphxpw”（重叠面片）。对于每个补丁，我都会做一些
处理。我现在要从图像中生成补丁这样，新图像必须与原始图像具有相同的形状。你能帮忙吗
让我来做。

这里](https://i.stack.imgur.com/kQkfw.png)
这是我提取补丁的代码

def patch_extract_3D(input,patch_shape,xstep=1,ystep=1,zstep=1):
    patches_3D = np.lib.stride_tricks.as_strided(input, ((input.shape[0] - patch_shape[0] + 1) / xstep, (input.shape[1] - patch_shape[1] + 1) / ystep,
                                                  (input.shape[2] - patch_shape[2] + 1) / zstep, patch_shape[0], patch_shape[1], patch_shape[2]),
                                                  (input.strides[0] * xstep, input.strides[1] * ystep,input.strides[2] * zstep, input.strides[0], input.strides[1],input.strides[2]))
    patches_3D= patches_3D.reshape(patches_3D.shape[0]*patches_3D.shape[1]*patches_3D.shape[2], patch_shape[0],patch_shape[1],patch_shape[2])
    return patches_3D

This is the processing the patches (just simple multiple with 2

for i in range(patches_3D.shape[0]):
    patches_3D[i]=patches_3D[i];
    patches_3D[i]=patches_3D[i]*2;

Now, what I need is from patches_3D, I want to reshape it to the original
image. Thanks

This is example code

patch_shape=[2, 2, 2]
input=np.arange(4*4*6).reshape(4,4,6)
patches_3D=patch_extract_3D(input,patch_shape)
print  patches_3D.shape
for i in range(patches_3D.shape[0]):
    patches_3D[i]=patches_3D[i]*2
print  patches_3D.shape

Answer 1

但是，这样做是相反的，因为你的补丁重叠，这只会
如果它们的值在重叠的地方一致，就要有明确的定义

def stuff_patches_3D(out_shape,patches,xstep=12,ystep=12,zstep=12):
    out = np.zeros(out_shape, patches.dtype)
    patch_shape = patches.shape[-3:]
    patches_6D = np.lib.stride_tricks.as_strided(out, ((out.shape[0] - patch_shape[0] + 1) // xstep, (out.shape[1] - patch_shape[1] + 1) // ystep,
                                                  (out.shape[2] - patch_shape[2] + 1) // zstep, patch_shape[0], patch_shape[1], patch_shape[2]),
                                                  (out.strides[0] * xstep, out.strides[1] * ystep,out.strides[2] * zstep, out.strides[0], out.strides[1],out.strides[2]))
    patches_6D[...] = patches.reshape(patches_6D.shape)
    return out

Update: here is a safer version that averages overlapping pixels:

def stuff_patches_3D(out_shape,patches,xstep=12,ystep=12,zstep=12):
    out = np.zeros(out_shape, patches.dtype)
    denom = np.zeros(out_shape, patches.dtype)
    patch_shape = patches.shape[-3:]
    patches_6D = np.lib.stride_tricks.as_strided(out, ((out.shape[0] - patch_shape[0] + 1) // xstep, (out.shape[1] - patch_shape[1] + 1) // ystep,
                                                  (out.shape[2] - patch_shape[2] + 1) // zstep, patch_shape[0], patch_shape[1], patch_shape[2]),
                                                  (out.strides[0] * xstep, out.strides[1] * ystep,out.strides[2] * zstep, out.strides[0], out.strides[1],out.strides[2]))
    denom_6D = np.lib.stride_tricks.as_strided(denom, ((denom.shape[0] - patch_shape[0] + 1) // xstep, (denom.shape[1] - patch_shape[1] + 1) // ystep,
                                                  (denom.shape[2] - patch_shape[2] + 1) // zstep, patch_shape[0], patch_shape[1], patch_shape[2]),
                                                  (denom.strides[0] * xstep, denom.strides[1] * ystep,denom.strides[2] * zstep, denom.strides[0], denom.strides[1],denom.strides[2]))
    np.add.at(patches_6D, tuple(x.ravel() for x in np.indices(patches_6D.shape)), patches.ravel())
    np.add.at(denom_6D, tuple(x.ravel() for x in np.indices(patches_6D.shape)), 1)
    return out/denom