在确定OpenCV用于非线性缩放因子的图像下采样的算法/实现时,我需要帮助。
我知道这个问题已经问过几次了,但是大多数答案似乎都不符合OpenCV的实现(例如,使用OpenCV时,此答案不正确:https://math.stackexchange.com/questions/48903/2d-array-downsampling-and-upsampling-using-bilinear-interpolation)。
最小问题表述:
我想使用双线性插值将4x4分辨率的图像降采样为3x3分辨率的图像。我对插值系数感兴趣。
python示例:
img = np.asarray([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
[ 9, 10, 11, 12],
[13, 14, 15, 16]]).astype(np.float32)
img_resized = cv2.resize(img, (3, 3), 0, 0, cv2.INTER_LINEAR).astype(np.float32)
print(img)
# [[ 1. 2. 3. 4.]
# [ 5. 6. 7. 8.]
# [ 9. 10. 11. 12.]
# [13. 14. 15. 16.]]
print(img_resized)
# [[ 1.8333333 3.1666667 4.5 ]
# [ 7.166667 8.5 9.833333 ]
# [12.5 13.833333 15.166666 ]]
插值系数:
经过反复试验,我弄清楚了OpenCV用于此特定情况的插值系数。
对于3x3图像的拐角点:
1.8333333 = 25/36 * 1 + 5/36 * 2 + 5/36 * 5 + 1/36 * 6
4.5000000 = 25/36 * 4 + 5/36 * 3 + 5/36 * 8 + 1/36 * 7
12.5000000 = 25/36 * 13 + 5/36 * 9 + 5/36 * 14 + 1/36 * 10
15.1666666 = 25/36 * 16 + 5/36 * 15 + 5/36 * 12 + 1/36 * 11
对于3x3图像的中间点:
8.5 = 1/4 * 6 + 1/4 * 7 + 1/4 * 10 + 1/4 * 11
对于3x3图像的其余4个点:
3.1666667 = 5/12 * 2 + 5/12 * 3 + 1/12 * 6 + 1/12 * 7
7.1666667 = 5/12 * 5 + 5/12 * 9 + 1/12 * 6 + 1/12 * 10
9.8333333 = 5/12 * 8 + 5/12 * 12 + 1/12 * 7 + 1/12 * 11
13.833333 = 5/12 * 14 + 5/12 * 15 + 1/12 * 10 + 1/12 * 11
问题:
有人可以帮助我理解这些插值系数吗?如何计算?我试图阅读cv :: resize()函数的源代码,但它对我没有太大帮助:S
答案 0 :(得分:1)
在研究了各种测试用例之后,我想我知道OpenCV如何选择采样点位置的答案。正如@ChrisLuengo在评论中指出的那样,OpenCV似乎没有在下采样之前应用低通滤波器,而是仅使用(双)线性插值。
(可能)解决方案:
让我们假设有一张5x5的图像,其像素位置在下图中用蓝色圆圈表示。现在,我们要将其降采样为3x3或4x4图像,并且需要在原始图像网格中找到新的降采样图像的采样位置。
似乎OpenCV对原始图像网格使用了1的像素距离,对新图像网格使用了(OLD_SIZE / NEW_SIZE)的像素距离,因此这里是5/3和5/4。而且,它将两个网格都对准中心点。因此,OpenCV的确定性采样算法可以如下所示:
可视化5x5至3x3 :
可视化5x5至4x4 :
示例代码(Python 2.7):
import numpy as np
import cv2
# 1. H_W is the height & width of the original image, using uniform H/W for this example
# resized_H_W is the height & width of the resized image, using uniform H/W for this example
H_W = 5
resized_H_W = 4
# 2. Create original image & Get OpenCV resized image:
img = np.zeros((H_W, H_W)).astype(np.float32)
counter = 1
for i in range(0, H_W):
for j in range(0, H_W):
img[i, j] = counter
counter += 1
img_resized_opencv = cv2.resize(img, (resized_H_W, resized_H_W), 0, 0, cv2.INTER_LINEAR).astype(np.float32)
# 3. Get own resized image:
img_resized_own = np.zeros((resized_H_W, resized_H_W)).astype(np.float32)
for i in range(0, resized_H_W):
for j in range(0, resized_H_W):
sample_x = (1.0 * H_W) / 2.0 - 0.50 + (i - (1.0 * resized_H_W - 1.0) / 2.0) * (1.0 * H_W) / (1.0 * resized_H_W)
sample_y = (1.0 * H_W) / 2.0 - 0.50 + (j - (1.0 * resized_H_W - 1.0) / 2.0) * (1.0 * H_W) / (1.0 * resized_H_W)
pixel_top_left = img[int(np.floor(sample_x)), int(np.floor(sample_y))]
pixel_top_right = img[int(np.floor(sample_x)), int(np.ceil(sample_y))]
pixel_bot_left = img[int(np.ceil(sample_x)), int(np.floor(sample_y))]
pixel_bot_right = img[int(np.ceil(sample_x)), int(np.ceil(sample_y))]
img_resized_own[i, j] = (1.0 - (sample_x - np.floor(sample_x))) * (1.0 - (sample_y - np.floor(sample_y))) * pixel_top_left + \
(1.0 - (sample_x - np.floor(sample_x))) * (sample_y - np.floor(sample_y)) * pixel_top_right + \
(sample_x - np.floor(sample_x)) * (1.0 - (sample_y - np.floor(sample_y))) * pixel_bot_left + \
(sample_x - np.floor(sample_x)) * (sample_y - np.floor(sample_y)) * pixel_bot_right
# 4. Print results:
print "\n"
print "Org. image: \n", img
print "\n"
print "Resized image (OpenCV): \n", img_resized_opencv
print "\n"
print "Resized image (own): \n", img_resized_own
print "\n"
print "MSE between OpenCV <-> Own: ", np.mean(np.square(img_resized_opencv - img_resized_own))
print "\n"
免责声明:
这只是我通过大约10个测试用例进行测试的理论。我并不是说这是100%正确。