前言

emmm…快开学了，手头只有红底证件照，但是学院要求要蓝底，这可咋办呢。懒得下ps了。自己撸起来吧。

方法一: lableme

lableme标注完后。得到一个json文件，然后将这种json文件转成掩码图.

# 代码来自 https://blog.csdn.net/hello_dear_you/article/details/120130155
import json
import numpy as np
import cv2
# read json file
with open("origin_json/mypic.json", "r") as f:
    data = f.read()
 
# convert str to json objs
data = json.loads(data)
 
# get the points 
points = data["shapes"][0]["points"]
points = np.array(points, dtype=np.int32)   # tips: points location must be int32
 
# read image to get shape
image = cv2.imread("origin_png/person.jpg")
 
# create a blank image
mask = np.zeros_like(image, dtype=np.uint8)
 
# fill the contour with 255
cv2.fillPoly(mask, [points], (255, 255, 255))
 
# save the mask 
cv2.imwrite("mask/person_mask.png", mask)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

大概是这样:

然后利用这个mask生成图片

# 参考自: https://www.jianshu.com/p/1961aa0c02ee
import cv2
import numpy as np


origin_png = 'origin_png/person.jpg'
# maskPath = 'mask/person_mask.png'
maskPath = 'mask/bmv2.png'
result_png = 'result_png/result_png.png'


maskImg = cv2.imread(maskPath)
img = cv2.imread(origin_png)
assert maskImg.shape == img.shape, 'maskImg.shape != origin_png.shape'

h, w = img.shape[0], img.shape[1]
print('图片宽度: {}, 高度: {}'.format(h, w))

rgb = (19,122,171)
bgr = (rgb[2], rgb[1], rgb[0])
# (B, G, R)
for i in range(h):
    for j in range(w):
        if (maskImg[i, j] == 0).all():
            img[i, j] = bgr
cv2.imwrite(result_png, img)
print('图片写入 {} 成功'.format(result_png))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

由于人长得一般，就不放图了…

缺点:
lableme标注时挺费力，并且难以避免人与背景边缘会有残留红色像素的情况。

方法二: 阈值

该方法通过比较像素的RGB与背景的RGB来区分是否为图像背景。

Opencv

import cv2
import numpy as np


def mean_square_loss(a_np, b_np):
    sl = np.square(a_np - b_np)
    return np.mean(sl)


def change_red2blue(origin_png, result_png):
    img = cv2.imread(origin_png)

    h, w = img.shape[0], img.shape[1]
    print('图片宽度: {}, 高度: {}'.format(h, w))

    origin_rgb = (168,36,32)  # 可以用浏览器啥的控制台工具提取出背景的rgb值
    origin_bgr = (origin_rgb[2], origin_rgb[1], origin_rgb[0])
    target_rgb = (19,122,171) # 蓝底RBG
    target_bgr = (target_rgb[2], target_rgb[1], target_rgb[0])

    for i in range(h):
        for j in range(w):
            # (B, G, R)
            if mean_square_loss(img[i, j], origin_bgr) < 50:
                img[i, j] = target_bgr 

    cv2.imwrite(result_png, img)
    print('图片写入 {} 成功'.format(result_png))


if __name__ == '__main__':
    # origin_png = 'result_png/result_png.png'
    origin_png = 'origin_png/person.jpg'
    result_png = 'result_png/result_refine.png'
    change_red2blue(origin_png, result_png)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35

结果人与背景边缘仍会存在红色像素残留

PIL

from torchvision.transforms.functional import to_tensor, to_pil_image
from PIL import Image
import torch
import time


def mean_square_loss(a_ts, b_ts):
    # print(a_ts.shape)
    # print(b_ts)
    sl = (a_ts - b_ts) ** 2
    return sl.sum()


def change_red2blue(origin_png, result_png):
    src = Image.open(origin_png)
    src = to_tensor(src)
    # print(src.shape)  # torch.Size([3, 800, 600])
    # channel: (R, G, B) / 255
    h, w = src.shape[1], src.shape[2]

    pha = torch.ones(h, w, 3)

    bg = torch.tensor([168,36,32]) / 255
    target_bg = torch.tensor([19,122,171]) / 255

    # C, H, W -> H, W, C
    src = src.permute(1, 2, 0)
    for i in range(h):
        for j in range(w):
            if mean_square_loss(src[i][j], bg) < 0.025: # 0.025是阈值，超参数
                pha[i][j] = torch.tensor([0.0, 0.0, 0.0])

    # H, W, C -> C, H, W
    src = src.permute(2, 0, 1)
    pha = pha.permute(2, 0, 1)
    com = pha * src + (1 - pha) * target_bg.view(3, 1, 1)
    to_pil_image(com).save(result_png)


if __name__ == '__main__':
    origin_png = 'origin_png/person.jpg'
    result_png = 'result_png/com.png'
    start_time = time.time()
    change_red2blue(origin_png, result_png)
    spend_time = round(time.time() - start_time, 2)
    print('生成成功，共花了 {} 秒'.format(spend_time))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46

该方法质量较好，但一张图片大概需要12秒。

方法四： Background MattingV2

Real-Time High-Resolution Background Matting
CVPR 2021 oral

论文：https://arxiv.org/abs/2012.07810
代码：https://github.com/PeterL1n/BackgroundMattingV2

github的readme.md有inference的colab链接，可以用那个跑

由于这篇论文是需要输入一张图片(例如有人存在的草地上)和背景图片的(如果草地啥的)， 然后模型会把人抠出来。

于是这里我需要生成一个背景图片。
首先我先借助firefox的颜色拾取器，得到十六进制，再用在线转换工具转成rgb。

然后生成一个背景图片。

import cv2
import numpy as np


image = cv2.imread("origin_png/person.jpg")
origin_rgb = (168,36,32)  # 可以用浏览器啥的控制台工具提取出背景的rgb值
origin_bgr = (origin_rgb[2], origin_rgb[1], origin_rgb[0])
image[:, :] = origin_bgr

cv2.imwrite("mask/bg.png", image)
1
2
3
4
5
6
7
8
9
10

然后输入给模型，然后换一下RGB。

然后就输出图片了。可以看到效果相当好。不愧是oral。

原论文可以实现发丝级效果