TibetanHWR 系列二：图像预处理——从红格纸扫描图到单字母图像

1
image, contours, hier = cv2.findContours(
2
    img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE
3
)
4
for c in contours:
5
    rect  = cv2.minAreaRect(c)   # 最小外接矩形
6
    box_  = cv2.boxPoints(rect)
7
    h = abs(box_[3, 1] - box_[1, 1])
8
    w = abs(box_[3, 0] - box_[1, 0])
9

10
    # 过滤掉不符合尺寸的轮廓
11
    if h > 3000 or w > 2200:
12
        continue
13
    if h < 2500 or w < 1500:
14
        continue
15

16
    box   = np.int0(cv2.boxPoints(rect))
17
    angle = rect[2]
18

19
    # 角度归一化到 [0, 45] 范围
20
    if abs(angle) > 45:
21
        angle = 90 - abs(angle)

1
import cv2
2
import numpy as np
3
import os
4
import replace
5

6

7
def separate_color_red(img):
8
    """提取图像中的红色区域（HSV 颜色提取）"""
9
    hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
10
    lower_hsv = np.array([0, 43, 46])
11
    high_hsv  = np.array([10, 255, 255])
12
    mask = cv2.inRange(hsv, lowerb=lower_hsv, upperb=high_hsv)
13
    print("颜色提取完成")
14
    return mask
15

16

17
def salt(img, n):
18
    """椒盐噪声滤波"""
19
    for k in range(n):
20
        i = int(np.random.random() * img.shape[1])
21
        j = int(np.random.random() * img.shape[0])
22
        if img.ndim == 2:
23
            img[j, i] = 255
24
        elif img.ndim == 3:
25
            img[j, i, 0] = 255
26
            img[j, i, 1] = 255
27
            img[j, i, 2] = 255
28
    return img
29

30

31
def cut(img, box, label, save_path):
32
    """按格子裁剪并保存图像"""
33
    x1, y1 = box[1]
34
    x2, y2 = box[3]
35
    img_cut = img[y1 + 10:y2 - 10, x1 + 10:x2 - 10]
36

37
    row = 8   # 行数
38
    col = 12  # 列数
39
    h, w = img_cut.shape[:2]
40
    cell_h = h // row
41
    cell_w = w // col
42

43
    os.makedirs(save_path, exist_ok=True)
44
    for r in range(row):
45
        for c in range(col):
46
            cell = img_cut[r * cell_h:(r + 1) * cell_h,
47
                           c * cell_w:(c + 1) * cell_w]
48
            filename = os.path.join(save_path, f"{label}_{r}_{c}.png")
49
            cv2.imwrite(filename, cell)
50
            replace.replace_red(filename)
51

52

53
def process(img_path, label, save_path):
54
    img = cv2.imread(img_path)
55
    # 1. 颜色提取
56
    mask = separate_color_red(img)
57
    # 2. 中值滤波
58
    mediu = cv2.medianBlur(mask, 19)
59
    # 3. 霍夫直线检测
60
    img_canny = cv2.Canny(mediu, 20, 250)
61
    lines = cv2.HoughLinesP(img_canny, 1, np.pi / 180, 120,
62
                            lines=4, minLineLength=50, maxLineGap=150)
63
    if lines is None:
64
        print(f"[警告] 未检测到直线：{img_path}")
65
        return
66
    lines1 = lines[:, 0, :]
67
    for x1, y1, x2, y2 in lines1:
68
        cv2.line(mediu, (x1, y1), (x2, y2), (255, 255, 255), 2)
69
    # 4. 轮廓检测
70
    _, contours, _ = cv2.findContours(
71
        mediu, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
72
    for c in contours:
73
        rect  = cv2.minAreaRect(c)
74
        box_  = cv2.boxPoints(rect)
75
        h = abs(box_[3, 1] - box_[1, 1])
76
        w = abs(box_[3, 0] - box_[1, 0])
77
        if h > 3000 or w > 2200:
78
            continue
79
        if h < 2500 or w < 1500:
80
            continue
81
        box   = np.int0(cv2.boxPoints(rect))
82
        angle = rect[2]
83
        if abs(angle) > 45:
84
            angle = 90 - abs(angle)
85
        # 5. 旋转矫正
86
        (ih, iw) = img.shape[:2]
87
        center = (iw // 2, ih // 2)
88
        M = cv2.getRotationMatrix2D(center, angle, 1)
89
        rotated = cv2.warpAffine(img, M, (iw, ih))
90
        # 6. 裁剪并保存
91
        cut(rotated, box, label, save_path)
92

93

94
if __name__ == "__main__":
95
    import sys
96
    img_path  = sys.argv[1]
97
    label     = sys.argv[2]
98
    save_path = sys.argv[3]
99
    process(img_path, label, save_path)