在扫描文档中拆分文本行

Question

我正在尝试找到一种方法来打破已自适应阈值化的扫描文档中的文本行拆分。现在，我将文档的像素值存储为从 0 到 255 的无符号整数，我取每行中像素的平均值，并根据像素值的平均值是否为大于 250，然后我取每行范围的中位数。但是，这种方法有时会失败，因为图像上可能会出现黑色斑点。

有没有更抗噪音的方法来完成这项任务？

编辑：这是一些代码。 “warped”是原始图像的名称，“cuts”是我要分割图像的位置。

warped = threshold_adaptive(warped, 250, offset = 10)
warped = warped.astype("uint8") * 255

# get areas where we can split image on whitespace to make OCR more accurate
color_level = np.array([np.sum(line) / len(line) for line in warped])
cuts = []
i = 0
while(i < len(color_level)):
    if color_level[i] > 250:
        begin = i
        while(color_level[i] > 250):
            i += 1
        cuts.append((i + begin)/2) # middle of the whitespace region
    else:
        i += 1

编辑 2：添加了示例图像

Answer 1

从您的输入图像中，您需要将文本设置为白色，并将背景设置为黑色

然后您需要计算账单的旋转角度。一个简单的方法是找到所有白点的minAreaRect (findNonZero)，你得到：

然后你可以旋转你的账单，让文字是水平的：

现在您可以计算水平投影 (reduce)。您可以取每行的平均值。在直方图上应用阈值th 以考虑图像中的一些噪声（这里我使用0，即没有噪声）。在直方图中，只有背景的行将具有值>0，文本行将具有值0。然后取直方图中每个连续白色 bin 序列的平均 bin 坐标。这将是您的线条的y 坐标：

这里是代码。它是用 C++ 编写的，但由于大部分工作都是使用 OpenCV 函数，所以它应该很容易转换为 Python。至少，您可以将此作为参考：

#include <opencv2/opencv.hpp>
using namespace cv;
using namespace std;

int main()
{
    // Read image
    Mat3b img = imread("path_to_image");

    // Binarize image. Text is white, background is black
    Mat1b bin;
    cvtColor(img, bin, COLOR_BGR2GRAY);
    bin = bin < 200;

    // Find all white pixels
    vector<Point> pts;
    findNonZero(bin, pts);

    // Get rotated rect of white pixels
    RotatedRect box = minAreaRect(pts);
    if (box.size.width > box.size.height)
    {
        swap(box.size.width, box.size.height);
        box.angle += 90.f;
    }

    Point2f vertices[4];
    box.points(vertices);

    for (int i = 0; i < 4; ++i)
    {
        line(img, vertices[i], vertices[(i + 1) % 4], Scalar(0, 255, 0));
    }

    // Rotate the image according to the found angle
    Mat1b rotated;
    Mat M = getRotationMatrix2D(box.center, box.angle, 1.0);
    warpAffine(bin, rotated, M, bin.size());

    // Compute horizontal projections
    Mat1f horProj;
    reduce(rotated, horProj, 1, CV_REDUCE_AVG);

    // Remove noise in histogram. White bins identify space lines, black bins identify text lines
    float th = 0;
    Mat1b hist = horProj <= th;

    // Get mean coordinate of white white pixels groups
    vector<int> ycoords;
    int y = 0;
    int count = 0;
    bool isSpace = false;
    for (int i = 0; i < rotated.rows; ++i)
    {
        if (!isSpace)
        {
            if (hist(i))
            {
                isSpace = true;
                count = 1;
                y = i;
            }
        }
        else
        {
            if (!hist(i))
            {
                isSpace = false;
                ycoords.push_back(y / count);
            }
            else
            {
                y += i;
                count++;
            }
        }
    }

    // Draw line as final result
    Mat3b result;
    cvtColor(rotated, result, COLOR_GRAY2BGR);
    for (int i = 0; i < ycoords.size(); ++i)
    {
        line(result, Point(0, ycoords[i]), Point(result.cols, ycoords[i]), Scalar(0, 255, 0));
    }

    return 0;
}

Answer 2

基本步骤 as @Miki,

1. 阅读源代码
2. 脱粒
3. 找到 minAreaRect
4. 被旋转矩阵扭曲
5. 查找并绘制上下界

---

而 Python 中的代码：

#!/usr/bin/python3
# 2018.01.16 01:11:49 CST
# 2018.01.16 01:55:01 CST
import cv2
import numpy as np

## (1) read
img = cv2.imread("img02.jpg")
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

## (2) threshold
th, threshed = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV|cv2.THRESH_OTSU)

## (3) minAreaRect on the nozeros
pts = cv2.findNonZero(threshed)
ret = cv2.minAreaRect(pts)

(cx,cy), (w,h), ang = ret
if w>h:
    w,h = h,w
    ang += 90

## (4) Find rotated matrix, do rotation
M = cv2.getRotationMatrix2D((cx,cy), ang, 1.0)
rotated = cv2.warpAffine(threshed, M, (img.shape[1], img.shape[0]))

## (5) find and draw the upper and lower boundary of each lines
hist = cv2.reduce(rotated,1, cv2.REDUCE_AVG).reshape(-1)

th = 2
H,W = img.shape[:2]
uppers = [y for y in range(H-1) if hist[y]<=th and hist[y+1]>th]
lowers = [y for y in range(H-1) if hist[y]>th and hist[y+1]<=th]

rotated = cv2.cvtColor(rotated, cv2.COLOR_GRAY2BGR)
for y in uppers:
    cv2.line(rotated, (0,y), (W, y), (255,0,0), 1)

for y in lowers:
    cv2.line(rotated, (0,y), (W, y), (0,255,0), 1)

cv2.imwrite("result.png", rotated)

---

最终结果：