在 Python 中使用双线性插值调整大小

Question

我正在尝试使用双线性插值手动实现调整图像大小。 到目前为止，我得到的是插值本身的功能，它似乎工作正常，但调整大小功能似乎只在角落正常工作：

def bilinear_interpolation(image, y, x):
height = len(image)
width = len(image[0])

x1 = min(math.floor(x), width - 1)
y1 = min(math.floor(y), height - 1)
x2 = min(math.ceil(x), width - 1)
y2 = min(math.ceil(y), height - 1)

a = image[y1][x1]
b = image[y2][x1]
c = image[y1][x2]
d = image[y2][x2]

new_pixel = a * (1 - x) * (1 - y)
new_pixel += b * y * (1 - x)
new_pixel += c * x * (1 - y)
new_pixel += d * x * y
return round(new_pixel)
def resize(image, new_height, new_width):

new_image = [[0 for _ in range(new_width)] for _ in range(new_height)]

for y in range(new_height):
    for x in range(new_width):
        x_ = (x / (new_width - 1)) * len(image[0])
        y_ = (y / (new_height - 1)) * len(image)

        new_image[y][x] = bilinear_interpolation(image, y_, x_)

return new_image

所以对于矩阵：

[[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]

我明白了

[[1, 4], [9, 12], [13, 16]]

而不是

[[1, 4], [7, 10], [13, 16]]

提前致谢。

Answer 1

计算x_和y_可以看我下面的answer。

公式为：

x_ = (x - x_scaled_center) * scale_x + x_orig_center
y_ = (y - y_scaled_center) * scale_y + y_orig_center

---

在双线性插值中，你错过了dx和dy的计算：

dx = x - x1
dy = y - y1

new_pixel = a * (1 - dx) * (1 - dy)
new_pixel += b * dy * (1 - dx)
new_pixel += c * dx * (1 - dy)
new_pixel += d * dx * dy

---

这是一个完整的代码示例（将结果与cv2.resize 进行比较）：

import cv2
import math
import numpy as np

def bilinear_interpolation(image, y, x):
    height = image.shape[0]
    width = image.shape[1]

    x1 = max(min(math.floor(x), width - 1), 0)
    y1 = max(min(math.floor(y), height - 1), 0)
    x2 = max(min(math.ceil(x), width - 1), 0)
    y2 = max(min(math.ceil(y), height - 1), 0)

    a = float(image[y1, x1])
    b = float(image[y2, x1])
    c = float(image[y1, x2])
    d = float(image[y2, x2])

    dx = x - x1
    dy = y - y1

    new_pixel = a * (1 - dx) * (1 - dy)
    new_pixel += b * dy * (1 - dx)
    new_pixel += c * dx * (1 - dy)
    new_pixel += d * dx * dy
    return round(new_pixel)


def resize(image, new_height, new_width):
    new_image = np.zeros((new_height, new_width), image.dtype)  # new_image = [[0 for _ in range(new_width)] for _ in range(new_height)]

    orig_height = image.shape[0]
    orig_width = image.shape[1]

    # Compute center column and center row
    x_orig_center = (orig_width-1) / 2
    y_orig_center = (orig_height-1) / 2

    # Compute center of resized image
    x_scaled_center = (new_width-1) / 2
    y_scaled_center = (new_height-1) / 2

    # Compute the scale in both axes
    scale_x = orig_width / new_width;
    scale_y = orig_height / new_height;

    for y in range(new_height):
        for x in range(new_width):
            x_ = (x - x_scaled_center) * scale_x + x_orig_center
            y_ = (y - y_scaled_center) * scale_y + y_orig_center

            new_image[y, x] = bilinear_interpolation(image, y_, x_)

    return new_image

img = cv2.imread('graf.png', cv2.IMREAD_GRAYSCALE)  # http://man.hubwiz.com/docset/OpenCV.docset/Contents/Resources/Documents/db/d70/tutorial_akaze_matching.html

new_width = 500
new_height = 400

resized_img = resize(img, new_height, new_width)

# Reference for testing
reference_resized_img = cv2.resize(img, (new_width, new_height), interpolation=cv2.INTER_LINEAR)

abs_diff = cv2.absdiff(reference_resized_img, resized_img)
print('max abs_diff = ' + str(abs_diff.max()))  # 1 gray level difference

cv2.imshow('img', img)
cv2.imshow('resized_img', resized_img)
cv2.imshow('reference_resized_img', reference_resized_img)
cv2.imshow('abs_diff*10', abs_diff*10)
cv2.waitKey()
cv2.destroyAllWindows()

---

img:

resized_img:

---

注意（对于其他用户）：

• 实施更多的是学术性而非实际性。
  如果您正在寻找一种有效的实施方式，请寻找其他地方。