python 图片常数代码 padding image constant pad with constant code

怼烎@ 2021-12-14 05:39 436阅读 0赞

简单来讲，计算机视觉需要将图片大小统一，这样才能输入到网络。

这个代码就是将2D或者3D的图片padding 到指定大小，边缘填充使用指定常数，

如果是3D，(512,512,3)最后一维不进行拓展

我加了注释，如果拓展的不能整除，那么左，上，也就是序号小的比序号大的差一维

rar = crop_to((5,5),np.array([[1,3,4,5],[2,3,3,3]]))
    print(rar)
    输出：
    [[21 21 21 21]
     [21 21 21 21]
     [ 1  3  4  5]
     [ 2  3  3  3]
     [21 21 21 21]]

import numpy as np
    
    
    def crop_to(tag_shape=(512,512), org_arr=np.array([0,0]),constant_value=21):
        """
        Padding a 2D array org_arr to a arbitrary tag_shape with a constant_value .
        If the horizon and vertical gap not divided by 2
        ,then left0>right0
        and up0>right0
        Only for 2D
        Args:
            tag_shape:1D array,int, target shape suppose bigger than origin array shape
            org_arr: 2D array prepared to be pad with 0 to shape equals to org_arr
            constant_value: int value for padding
    
        Returns:2D padded array
    
        """
        org_shape = org_arr.shape
        assert len(org_arr.shape) == 2 or len(org_shape)==3
        assert len(tag_shape) == len(org_shape)
    
        assert org_shape[0] <= tag_shape[0]
        assert org_shape[1] <= tag_shape[1]
        horiz = tag_shape[0] - org_shape[0]
        vertc = tag_shape[1] - org_shape[1]
        if horiz % 2 == 0:
            left = int(horiz / 2)
            right = left
        else:
            left = int(horiz / 2) + 1
            right = left - 1
        if vertc % 2 == 0:
            up = int(vertc / 2)
            down = up
        else:
            up = int(vertc / 2) + 1
            down = up -1
        if len(org_shape)==2:
            tar_arr = np.pad(org_arr, ((left, right), (up, down)),mode='constant',constant_values=constant_value)
        else:
            tar_arr = np.pad(org_arr, ((left, right), (up, down),(0,0)), mode='constant', constant_values=constant_value)
    
        return tar_arr
    # rar = crop_to((5,5),np.array([[1,3,4,5],[2,3,3,3]]))
    # print(rar)

以及相应的代码还原回去：

def unpadding(tag_shape=(512,512), org_arr=np.array([0,0])):
        org_shape = org_arr.shape
        assert len(org_arr.shape) == 2 or len(org_shape) == 3
        assert len(tag_shape) == len(org_shape)
    
        assert org_shape[0] >= tag_shape[0]
        assert org_shape[1] >= tag_shape[1]
        horiz = org_shape[0]-tag_shape[0]
        vertc = org_shape[1]-tag_shape[1]
        if horiz % 2 == 0:
            left = int(horiz / 2)
            right = org_shape[0]-left
        else:
            left = int(horiz / 2) + 1
            right = org_shape[0]-left + 1
        if vertc % 2 == 0:
            up = int(vertc / 2)
            down = org_shape[1]-up
        else:
            up = int(vertc / 2) + 1
            down = org_shape[1]-up +1
        if len(org_shape) == 2:
            tar_arr = org_arr[left:right,up:down]
            #tar_arr = np.pad(org_arr, ((left, right), (up, down)), mode='constant', constant_values=constant_value)
        else:
            tar_arr = org_arr[left:right,up,down]
            #tar_arr = np.pad(org_arr, ((left, right), (up, down), (0, 0)), mode='constant', constant_values=constant_value)
        return tar_arr

使用样例：

rar = crop_to((5,5),np.array([[1,3,4,5],[2,3,3,3]]))
    print(rar)
    up = unpadding((2,4),rar)
    print(up)
    输出：
    [[21 21 21 21 21]
     [21 21 21 21 21]
     [21  1  3  4  5]
     [21  2  3  3  3]
     [21 21 21 21 21]]
    [[1 3 4 5]
     [2 3 3 3]]

opencv是有padding代码，但我找不到handy的教程，所以直接写了一个，如果有更好的建议，欢迎评论。