OpenCV实现页脚翻折效果

最新推荐文章于 2022-08-14 15:26:02 发布

原创最新推荐文章于 2022-08-14 15:26:02 发布 · 2.6k 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#opencv #人工智能 #计算机视觉

OpenCV 专栏收录该内容

1 篇文章

订阅专栏

本文介绍了一种在文档图像上实现翻折效果的方法，通过坐标变换实现文档的局部翻折，以此来增加图像数据集的变化性和真实性。文章详细介绍了翻折效果的数学原理和具体的实现代码。

由于需要对文档图像数据进行增强，需要对完好的文档图像加个页脚翻折效果，对文档进行折旧化。

效果(左边为效果图，右边为原图)：

思路

针对某个顶点，先确定一个矩阵，然后将上三角的矩阵的值赋值到下三角，同时将上三角的值置为白色(255)。其中最重要的是确定对角线和一个点关于这条对角线对称的点。

关于直线对称点的公式：

其中A, B, C为一般式方程的系数。通过这样就可以确定对称点的坐标。

代码如下：

import cv2
import numpy as np
import random
import json


def get_sym(a, b, coe):
    # A, B, C = 1, -1, -5
    A, B, C = coe
    # a, b = 1, -5
    # 对称点公式 https://zhidao.baidu.com/question/1691896311611343348.html
    x = -(2 * A * B * b + (A ** 2 - B ** 2)
          * a + 2 * A * C) / (A ** 2 + B ** 2)
    y = -(2 * A * B * a + (B ** 2 - A ** 2)
          * b + 2 * B * C) / (A ** 2 + B ** 2)
    return int(x), int(-y)


def fold_top_left(img, mask, point):
    pad = random.randint(50, 500)
    coe = (2, -1, -pad)
    A = random.randint(1, 5)
    B = random.randint(-5, -1)
    C = -pad
    coe = (A, B, C)
    A, B, C = coe

    w, h = int(point[0]), int(point[1])
    img_part = img[h:h + pad, w: w+pad]

    for i in range(pad):
        for j in range(pad):
            if A * i - B * j + C < 0:
                i_, j_ = get_sym(i, -j, coe)
                img_part[i_-1][j_-1] = img_part[i][j]
                img_part[i][j] = np.array([255, 255, 255])
                mask[h+i][w+j] = 255

    img[h:h + pad, w: w+pad] = img_part

    return img, mask


def fold_top_right(img, mask, point):
    coe = (5, 2, 0)
    A, B, C = coe

    A = random.randint(1, 10)
    B = random.randint(1, 5)
    while B > A:
        B = random.randint(1, 5)
    C = 0
    coe = (A, B, C)

    print(coe)

    w, h = int(point[0]), int(point[1])

    pad = random.randint(50, 500)
    img_part = img[h:h+pad, w-pad:w]
    img_mask = np.zeros((img.shape[0], img.shape[1]), dtype=np.uint8)

    for i in range(pad):
        for j in range(pad):
            if A * i - B * j + C < 0:
                i_, j_ = get_sym(i, -j, coe)
                # print(i_, j_)
                img_part[i_-1][j_-1] = img_part[i][j]
                img_part[i][j] = np.array([255, 255, 255])
                mask[h+i][w-pad+j] = 255


    img[h:h+pad, w-pad:w] = img_part

    return img, mask


def fold_bottom_right(img, mask, point):
    pad = random.randint(50, 500)
    w, h = int(point[0]), int(point[1])

    pad = random.randint(50, 500)
    img_part = img[h-pad:h, w-pad:w]

    for i in range(pad):
        for j in range(pad):
            if i+j > pad:
                img_part[pad - 1 - j][pad - 1 - i] = img_part[i][j]
                img_part[i][j] = np.array([255, 255, 255])
                mask[h-pad+i][w-pad+j] = 255


    img[h-pad:h, w-pad:w] = img_part

    return img, mask


def fold_bottom_left(img, mask, point):
    # coe = (2, 5, 0)
    # A, B, C = coe

    A = random.randint(1, 5)
    B = random.randint(1, 10)
    while B < A:
        B = random.randint(1, 10)
    C = 0
    coe = (A, B, C)

    w, h = int(point[0]), int(point[1])

    pad = random.randint(50, 500)
    img_part = img[h-pad:h, w: w+pad]
    print(pad)
    for i in range(pad):
        for j in range(pad):
            if A * i - B * j + C > 0:
                i_, j_ = get_sym(i, -j, coe)
                img_part[i_][j_] = img_part[i][j]
                img_part[i][j] = np.array([255, 255, 255])
                mask[h-pad+i][w+j] = 255
                

    img[h-pad:h, w: w+pad] = img_part

    return img, mask



def generate_flip(img, mask, points):

    top_left = points[0]
    top_right = points[1]
    bottom_right = points[2]
    bottom_left = points[3]

    # 随机对四个角进行翻折
    tmp = random.randint(1, 100)
    if tmp % 2 == 0:
        img, mask = fold_top_left(img, mask, top_left)
    tmp = random.randint(1, 100)
    if tmp % 2 == 0:
        img, mask = fold_top_right(img, mask, top_right)
    tmp = random.randint(1, 100)
    if tmp % 2 == 0:
        img, mask = fold_bottom_right(img, mask, bottom_right)
    tmp = random.randint(1, 100)
    if tmp % 2 == 0:
        img, mask = fold_bottom_left(img, mask, bottom_left)

    cv2.imwrite('flipped.jpg', img)
    cv2.imwrite('mask.jpg', mask)
    return img, mask



if __name__ == '__main__':
    img = cv2.imread('4_bf.jpg')

    js_data = json.loads(open('4_bf.json').read())
    points = js_data['shapes'][0]['points']
    points = get_points(points)
    
    mask = np.zeros((img.shape[0], img.shape[1]), dtype=np.uint8)
    img = generate_flip(img, mask, points)