python 倾斜图片纠正 python 图片矫正

图像矫正就是通过一些已知的参考点，即无失真图象的某些象素点和畸变图象相应象素的坐标间对应关系，拟合出映射关系中的未知系数，并作为恢复其它象素的基础。

1、矫正原理

在进行图片矫正时，有些图片具有小角度的倾斜（±45°以内），导致传入后续识别分类的模型时产生误差。可以利用文本图像具有行间空白的特性，对待检测图像进行角度旋转遍历，并同时进行水平方向像素值投影，当文本方向水平时，投影所得的0值最多。从而达到所要矫正的目的。

2、边缘投影

在参考点足够的情况下，推算全图变形函数，比如：当水平线倾斜时投影值比较小，但是当水平线正常时其在水平面上的投影是最大的。根据这个进行参考，倾斜的物体在水平方向上的投影只要经过旋转处理就可以达到最大，此时的图片就是正常。

3、基于边缘投影法的矫正过程

（1）对检测图像进行角度旋转遍历

（2）进行水平方向像素值投影

（3）找到投影所得的0值最多的方向

（4）对该方向进行矫正

用到的工具：opencv、numpy、matplotlib

首先获取图片的宽度以及高度

def

其次，寻找当前图片文本的旋转角度，在±90度之间

def findangle(_image):
    toWidth = _image.shape[1]//2 #500
    minCenterDistance = toWidth/20 #10
    angleThres = 45

    image = _image.copy()
    h, w = image.shape[0:2]
    if w > h:
        maskW = toWidth
        maskH = int(toWidth / w * h)
    else:
        maskH = toWidth
        maskW = int(toWidth / h * w)

接着，使用黑色填充图片的其他区域

# 使用黑色填充图片区域
    swapImage = cv2.resize(image, (maskW, maskH))
    grayImage = cv2.cvtColor(swapImage, cv2.COLOR_BGR2GRAY)
    gaussianBlurImage = cv2.GaussianBlur(grayImage, (3, 3), 0, 0)
    histImage = cv2.equalizeHist(~gaussianBlurImage)
    binaryImage = cv2.adaptiveThreshold(histImage, 1, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, 
                                        cv2.THRESH_BINARY, 15, -2)

然后，遍历角度时计算的关键点数量

pointsNum = np.sum(binaryImage!=0)//2

紧接着使用连接组件寻找并删除边框线条，计算每个连通区域坐上右下点的索引坐标与其质心的距离，距离大的即为线条

connectivity = 8
    num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(binaryImage, connectivity, cv2.CV_8U)
    labels = np.array(labels)
    maxnum = [(i, stats[i][-1], centroids[i]) for i in range(len(stats))]
    maxnum = sorted(maxnum, key=lambda s: s[1], reverse=True)
    if len(maxnum) <= 1:
        return 0
    for i, (label, count, centroid) in enumerate(maxnum[1:]):
        cood = np.array(np.where(labels == label))
        distance1 = np.linalg.norm(cood[:,0]-centroid[::-1])
        distance2 = np.linalg.norm(cood[:,-1]-centroid[::-1])
        if distance1 > minCenterDistance or distance2 > minCenterDistance:
            binaryImage[labels == label] = 0
        else:
            break

最后，横向统计非零元素个数 越少则说明姿态越正

# 横向统计非零元素个数 越少则说明姿态越正
        zeroCount = np.sum(hist > toWidth/50)
        if zeroCount <= minCount or minCount == -1:
            minCount = zeroCount
            minRotate = rotate
    return minRotate

完整代码

import cv2
import numpy as np
from matplotlib import pyplot as plt

def rotate_bound(image, angle):# 获取图片的宽高
    (h, w) = image.shape[:2]
    (cX, cY) = (w // 2, h // 2)
    M = cv2.getRotationMatrix2D((cX, cY), angle, 1.0)
    img = cv2.warpAffine(image, M, (w, h))
    return img

def rotate_points(points, angle, cX, cY):
    M = cv2.getRotationMatrix2D((cX, cY), angle, 1.0).astype(np.float16)
    a = M[:, :2]
    b = M[:, 2:]
    b = np.reshape(b, newshape=(1, 2))
    a = np.transpose(a)
    points = np.dot(points, a) + b
    points = points.astype(np.int)
    return points

def findangle(_image):
    toWidth = _image.shape[1]//2 #500
    minCenterDistance = toWidth/20 #10
    angleThres = 45

    image = _image.copy()
    h, w = image.shape[0:2]
    if w > h:
        maskW = toWidth
        maskH = int(toWidth / w * h)
    else:
        maskH = toWidth
        maskW = int(toWidth / h * w)
    # 使用黑色填充图片区域
    swapImage = cv2.resize(image, (maskW, maskH))
    grayImage = cv2.cvtColor(swapImage, cv2.COLOR_BGR2GRAY)
    gaussianBlurImage = cv2.GaussianBlur(grayImage, (3, 3), 0, 0)
    histImage = cv2.equalizeHist(~gaussianBlurImage)
    binaryImage = cv2.adaptiveThreshold(histImage, 1, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 15, -2)
    
    pointsNum = np.sum(binaryImage!=0)//2

    connectivity = 8
    num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(binaryImage, connectivity, cv2.CV_8U)
    labels = np.array(labels)
    maxnum = [(i, stats[i][-1], centroids[i]) for i in range(len(stats))]
    maxnum = sorted(maxnum, key=lambda s: s[1], reverse=True)
    if len(maxnum) <= 1:
        return 0
    for i, (label, count, centroid) in enumerate(maxnum[1:]):
        cood = np.array(np.where(labels == label))
        distance1 = np.linalg.norm(cood[:,0]-centroid[::-1])
        distance2 = np.linalg.norm(cood[:,-1]-centroid[::-1])
        if distance1 > minCenterDistance or distance2 > minCenterDistance:
            binaryImage[labels == label] = 0
        else:
            break

    minRotate = 0
    minCount = -1
    (cX, cY) = (maskW // 2, maskH // 2)
    points = np.column_stack(np.where(binaryImage > 0))[:pointsNum].astype(np.int16)
    for rotate in range(-angleThres, angleThres):
        rotatePoints = rotate_points(points, rotate, cX, cY)
        rotatePoints = np.clip(rotatePoints[:,0], 0, maskH-1)
        hist, bins = np.histogram(rotatePoints, maskH, [0, maskH])
        # 横向统计非零元素个数 越少则说明姿态越正
        zeroCount = np.sum(hist > toWidth/50)
        if zeroCount <= minCount or minCount == -1:
            minCount = zeroCount
            minRotate = rotate
    return minRotate

Path = '003.jpg'
cv_img = cv2.imdecode(np.fromfile(Path, dtype=np.uint8), -1)
cv_img = cv2.cvtColor(cv_img, cv2.COLOR_RGB2BGR)

for agl in range(-30, 30):
    img = cv_img.copy()
    img = rotate_bound(img, agl)
    angle = findangle(img)
    img = rotate_bound(img, -angle)
    angle = findangle(img)
    cv2.imshow('after', img)
    key = cv2.waitKey(1) & 0xFF
    # 按'q'健退出循环
    if key == ord('q'):
        break
cv2.destroyAllWindows()
        
plt.imshow(img)
plt.show()

与原图效果进行比较

其他图片的矫正效果

相比较于用直线勘测的方法对图像进行矫正