将3D图像堆叠到4D

Question

我正在尝试堆叠3D图像以具有4D阵列。我的代码为：

def stack():
   x=None
   dim=(299,299)
   for file in os.listdir(path_to_folder):
        if file.endswith('.jpg'):
            img = cv2.imread(path_to_folder+ file)
            image_a = cv2.resize(img,dim, interpolation = cv2.INTER_AREA)

            img2 = cv2.imread(path_to_folder+ file)
            image_p = cv2.resize(img2,dim, interpolation = cv2.INTER_AREA)

            img3 = cv2.imread(path_to_folder+ file)
            image_n = cv2.resize(img3,dim, interpolation = cv2.INTER_AREA)

            if (x is None):
                x=[(image_a),(image_p),(image_n)]
            else:
                x[0]=np.stack((x[0], (image_a)))
                x[1]= np.stack((x[1],(image_p)))
                x[2]=np.stack((x[2],(image_n)))

    return x

我期望形状为：

stack=stack()
stack[0].shape
>>out: (5,299,299,3)

len(stack)
>>out: 3

但是我得到了(1495,299,3)。注意：为了只关注实际问题，我将所有3个图像文件都保持不变。我已经从here中使用了[[generate_triplets函数，但就我而言，我的图像是从文件夹中读取的。

Answer 1

有很多方法可以做到这一点，我将探索其中的一些方法。首先，np.stack要求所有数组的形状相同。您不能像这样反复调用np.stack。一个简单的解决方法是将所有图像存储在一个列表中，然后在最后调用堆栈。让我整理一些伪代码：

import numpy as np

def stack():
    x = [[],[],[]]
    dim = (299,299)
    for i in range(5):
        img_a = np.random.randn(dim[0],dim[1],3)
        img_p = np.random.randn(dim[0],dim[1],3)
        img_n = np.random.randn(dim[0],dim[1],3)

        x[0].append(img_a)
        x[1].append(img_p)
        x[2].append(img_n)

    x = [np.stack(im) for im in x]
    return x

stack = stack()
print(out[0].shape)

输出：
(5, 299, 299, 3)
3

但是，如果出于某种原因要在每次迭代中进行堆栈，您仍然可以使用vstack进行操作。您只需要通过重塑使每个3d图像成为4d图像即可。
import numpy as np

def stack():
    x = None
    dim = (299,299)
    for i in range(5):
        img_a = np.random.randn(dim[0],dim[1],3)
        img_p = np.random.randn(dim[0],dim[1],3)
        img_n = np.random.randn(dim[0],dim[1],3)

        if (x is None):
            x=[img_a,img_p,img_n]
        else:
            s = (-1,dim[0],dim[1],3)
            x[0]=np.vstack((x[0].reshape(s), img_a.reshape(s)))
            x[1]=np.vstack((x[1].reshape(s), img_p.reshape(s)))
            x[2]=np.vstack((x[2].reshape(s), img_n.reshape(s)))
    return x

stack = stack()
print(stack[0].shape)
print(len(stack))

输出：
(5, 299, 299, 3)
3

Answer 2

修改@Mercury的功能：

In [398]: def stack(): 
     ...:     alist = [] 
     ...:     dim = (299,299) 
     ...:     for i in range(5): 
     ...:         img_a = np.random.randn(dim[0],dim[1],3) 
     ...:         img_p = np.random.randn(dim[0],dim[1],3) 
     ...:         img_n = np.random.randn(dim[0],dim[1],3) 
     ...:         alist.append([img_a, img_p, img_n]) 
     ...:     return np.array(alist) 

In [399]: stack().shape                                                                        
Out[399]: (5, 3, 299, 299, 3)

alist是包含（299,299,3）数组的嵌套（5,3）列表。变成数组变成5d。
我们可以换位，stack().transpose(1,0,2,3,4)。但是np.stack是连接的版本，可让我们指定new轴：
In [400]: def stack(): 
     ...:     alist = [] 
     ...:     dim = (299,299) 
     ...:     for i in range(5): 
     ...:         img_a = np.random.randn(dim[0],dim[1],3) 
     ...:         img_p = np.random.randn(dim[0],dim[1],3) 
     ...:         img_n = np.random.randn(dim[0],dim[1],3) 
     ...:         alist.append([img_a, img_p, img_n]) 
     ...:     return np.stack(alist, axis=1) 
     ...:                                                                                      
In [401]: stack().shape                                                                        
Out[401]: (3, 5, 299, 299, 3)