Opencv 答题卡识别

Opencv 答题卡识别

大致步骤:

1 读取图像

2 预处理

​ 2.1 灰度转换

​ 2.2 高斯滤波

​ 2.3 边缘检测

3 轮廓检测

4 透视变换

5 阈值化

6 再次轮廓检测,从结果中筛选出符合的选项

7 遍历每一组的轮廓(即每一行的)

8 遍历每一组的每一个轮廓,通过掩码计算结果,取最大的那个选项的下标(下标即代表选项A,B,C,D,E)

9 以此类推,遍历每一组

10 绘图,输出结果


image

图像预处理

#预处理1
def preprocess(img):
    # 灰度转换
    imgGray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    #高斯滤波
    imgBlur = cv2.GaussianBlur(imgGray, (5, 5), 0)

    cv2.imshow('imgBlur', imgBlur)
    cv2.waitKey()
    #边缘检测
    imgEdge = cv2.Canny(imgBlur, 75, 200)
    cv2.imshow('Canny', imgEdge)
    cv2.waitKey()
	#返回预处理的结果
    return imgEdge

获取轮廓

def getCnts(img):
    cnts = cv2.findContours(img.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[1]
    docCnts = None
    # 确保有轮廓
    if len(cnts) > 0:
        cnts = sorted(cnts, key=cv2.contourArea, reverse=True)
        for c in cnts:
            # 轮廓近似
            peri = cv2.arcLength(c, True)
            approx = cv2.approxPolyDP(c, 0.02 * peri, True)

            if len(approx) == 4:
                docCnts = approx
                break

        return docCnts

透视变换

image

如图所示,通过这种计算方法可以得到A,B,C,D四个点的坐标。

而透视变换则需要图像的原始坐标和变换后的坐标,变换后的坐标可以设置为

[0, 0], [w - 1, 0],[0, h - 1], [w - 1, h - 1]

因此,透视变换的两个函数的代码:

一是获取原始坐标,二是进行透视变换,返回透视变换的结果

def four_point_transfrom(img, pts):
    newPoints = []
    sumPoints = []
    subPoints = []
    for x, y in pts:
        sumPoints.append(x + y)
        subPoints.append(x - y)

    # get index
    newPoints.append(pts[(np.argmin(sumPoints))])#0-A

    newPoints.append(pts[(np.argmax(subPoints))])#1-B
    newPoints.append(pts[(np.argmin(subPoints))])#2-C

    newPoints.append(pts[(np.argmax(sumPoints))])#3-D
	
    return newPoints

def getWrap(img, wraped, w, h):
    src = np.array(wraped, dtype="float32")
    dst = np.array([[0, 0],
                    [w - 1, 0],
                    [0, h - 1],
                    [w - 1, h - 1]], dtype="float32")

    # src and dst must be type of float32
    M = cv2.getPerspectiveTransform(src, dst)

    imgRes = cv2.warpPerspective(img, M, (w, h))

    return imgRes

image

阈值化,再次轮廓检测

# threshold
imgWrapGray = cv2.cvtColor(imgWrap, cv2.COLOR_BGR2GRAY)
imgThreshold = cv2.threshold(imgWrapGray, 20, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1]
cv2.imshow("Wraped and threshold", imgThreshold)
cv2.waitKey()

# find Contours2

draw_cnts = imgWrap.copy()
thresh_cnts = cv2.findContours(imgThreshold.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[1]
cv2.drawContours(draw_cnts, thresh_cnts, -1, (0, 0, 255), 3)

image

过滤圆形轮廓

# 遍历所有轮廓
for c in thresh_cnts:
    # 计算比例和大小
    (x, y, w, h) = cv2.boundingRect(c)
    ar = w / float(h)


    # 根据实际情况指定标准
    if w >= 50 and h >= 40 and ar >= 0.9 and ar <= 1.6:
        print('w=', w)
        print('h=', h)
        print('---------------')#打印出来看效果更直观
        questionCnts.append(c)

对圆形轮廓继续遍历

对于同一个题目,x不同,y相同

对于不同题目的同一选项,x相同,y不同

因此,在排序时,先对每组进行排序,即,按y排序

questionCnts = sort_contours(questionCnts, method="top-to-bottom")[0]

其中

def sort_contours(cnts, method="left-to-right"):
    reverse = False
    i = 0
    if method == "right-to-left" or method == "bottom-to-top":
        reverse = True
    if method == "top-to-bottom" or method == "bottom-to-top":
        i = 1
    boundingBoxes = [cv2.boundingRect(c) for c in cnts]
    (cnts, boundingBoxes) = zip(*sorted(zip(cnts, boundingBoxes),
                                        key=lambda b: b[1][i], reverse=reverse))
    return cnts, boundingBoxes

每排有5个选项,继续遍历每个选项

for (q, i) in enumerate(np.arange(0, len(questionCnts), 5)):
    # 排序-按x坐标,每次排一组,一组为5个
    cnts = sort_contours(questionCnts[i:i + 5])[0]
    bubbled = None
    # 遍历每一组结果
    for (j, c) in enumerate(cnts):
        # 使用mask来判断结果
        mask = np.zeros(imgThreshold.shape, dtype="uint8")
        cv2.drawContours(mask, [c], -1, 255, -1)  # -1表示填充,[c]表示当前选项的位置

        # 通过计算非零点数量来算是否选择这个答案
        mask = cv2.bitwise_and(imgThreshold, imgThreshold, mask=mask)#相当于只保留了白色的部分
        total = cv2.countNonZero(mask)
        print('total=',total)
        #判断一下,如果他涂了两个选项,应该另外处理
        if total>1000:
            count += 1
        if count< 2:
            # 通过阈值判断,如果是第一次/当前值比上一次的大,则bubbled记录下来,直到最后bubbled就是一组中值最大的那个选项
            if bubbled is None or total > bubbled[0]:
                bubbled = (total, j)
        else:
            print('[Warning!]题目[%d]选择了多个选项'.format(j))
    print('-------------------')
    count=0

image
image

计算正确率,输出信息

这里答题卡是5行5列,利用计算出的5列的涂黑的值与真实答案的位置是否对应,对比正确答案,汇出图形。

       # 对比正确答案
    color = (0, 0, 255)
    k = ANSWER_KEY[q]

    # 判断正确,bubbled=(656,1),其中第二个依次表示A,B,C,D,E,
    if k == bubbled[1]:
        color = (0, 255, 0)
        correct += 1

    # 绘图
    cv2.drawContours(imgWrap, [cnts[k]], -1, color, 3)
    score = (correct / 5.0) * 100
    
#计算正确率,输出信息
print("[INFO] score: {:.2f}%".format(score))
cv2.putText(imgWrap, "{:.2f}%".format(score), (10, 30),
            cv2.FONT_HERSHEY_SIMPLEX, 0.9, (0, 0, 255), 2)
cv2.imshow("Original", imgOriginal)
cv2.imshow("Exam", imgWrap)
cv2.waitKey()

完整代码

import cv2
import numpy as np

# 正确答案
ANSWER_KEY = {0: 1, 1: 4, 2: 0, 3: 3, 4: 1}
def preprocess(img):
    # RGB2GRAY
    imgGray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    imgBlur = cv2.GaussianBlur(imgGray, (5, 5), 0)

    cv2.imshow('imgBlur', imgBlur)
    cv2.waitKey()
    imgEdge = cv2.Canny(imgBlur, 75, 200)
    cv2.imshow('Canny', imgEdge)
    cv2.waitKey()

    return imgEdge


def getCnts(img):
    cnts = cv2.findContours(img.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[1]
    docCnts = None
    # make sure detected
    if len(cnts) > 0:
        cnts = sorted(cnts, key=cv2.contourArea, reverse=True)
        for c in cnts:
            # appro
            peri = cv2.arcLength(c, True)
            approx = cv2.approxPolyDP(c, 0.02 * peri, True)

            if len(approx) == 4:
                docCnts = approx
                break

        return docCnts


def four_point_transfrom(img, pts):
    newPoints = []
    sumPoints = []
    subPoints = []
    for x, y in pts:
        sumPoints.append(x + y)
        subPoints.append(x - y)

    # get index
    newPoints.append(pts[(np.argmin(sumPoints))])

    newPoints.append(pts[(np.argmax(subPoints))])
    newPoints.append(pts[(np.argmin(subPoints))])

    newPoints.append(pts[(np.argmax(sumPoints))])

    return newPoints


def getWrap(img, wraped, w, h):
    src = np.array(wraped, dtype="float32")
    dst = np.array([[0, 0],
                    [w - 1, 0],
                    [0, h - 1],
                    [w - 1, h - 1]], dtype="float32")

    # src and dst must be type of float32
    M = cv2.getPerspectiveTransform(src, dst)

    imgRes = cv2.warpPerspective(img, M, (w, h))

    return imgRes


def sort_contours(cnts, method="left-to-right"):
    reverse = False
    i = 0
    if method == "right-to-left" or method == "bottom-to-top":
        reverse = True
    if method == "top-to-bottom" or method == "bottom-to-top":
        i = 1
    boundingBoxes = [cv2.boundingRect(c) for c in cnts]
    (cnts, boundingBoxes) = zip(*sorted(zip(cnts, boundingBoxes),
                                        key=lambda b: b[1][i], reverse=reverse))
    return cnts, boundingBoxes


# read img
imgOriginal = cv2.imread('test_02.png')
imgOriginal = cv2.resize(imgOriginal, (500, 600))
w = imgOriginal.shape[0]
h = imgOriginal.shape[1]
cv2.imshow('imgOriginal', imgOriginal)
cv2.waitKey()
drawImg = imgOriginal.copy()

# preprocess
imgPre = preprocess(drawImg)

# contours
cur_cnts = getCnts(imgPre)

# imgPerspective
wraped = four_point_transfrom(imgPre, cur_cnts.reshape(4, 2))
imgWrap = getWrap(imgOriginal, wraped, w, h)
cv2.imshow('imgWrap', imgWrap)
cv2.waitKey()

# threshold
imgWrapGray = cv2.cvtColor(imgWrap, cv2.COLOR_BGR2GRAY)
imgThreshold = cv2.threshold(imgWrapGray, 20, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1]
cv2.imshow("Wraped and threshold", imgThreshold)
cv2.waitKey()

# find Contours2

draw_cnts = imgWrap.copy()
thresh_cnts = cv2.findContours(imgThreshold.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)[1]
cv2.drawContours(draw_cnts, thresh_cnts, -1, (0, 0, 255), 3)

cv2.imshow("thresh_cnts", draw_cnts)
cv2.waitKey()

questionCnts = []

# 过滤出圆形轮廓
# 遍历
for c in thresh_cnts:
    # 计算比例和大小
    (x, y, w, h) = cv2.boundingRect(c)
    ar = w / float(h)


    # 根据实际情况指定标准
    if w >= 50 and h >= 40 and ar >= 0.9 and ar <= 1.6:
        print('w=', w)# 1.(24,15)
        print('h=', h)
        print('---------------')
        questionCnts.append(c)

# 对于同一个题,x相同,y不同
# 对于不同题,x不同,y相同

questionCnts = sort_contours(questionCnts, method="top-to-bottom")[0]
correct = 0
count = 0

# 每排有5个选项
for (q, i) in enumerate(np.arange(0, len(questionCnts), 5)):
    # 排序-按y坐标,每次排一组,一组为5个
    cnts = sort_contours(questionCnts[i:i + 5])[0]
    bubbled = None
    # 遍历每一组结果
    for (j, c) in enumerate(cnts):
        # 使用mask来判断结果
        mask = np.zeros(imgThreshold.shape, dtype="uint8")
        cv2.drawContours(mask, [c], -1, 255, -1)  # -1表示填充,[c]表示当前选项的位置

        # 通过计算非零点数量来算是否选择这个答案
        mask = cv2.bitwise_and(imgThreshold, imgThreshold, mask=mask)#相当于只保留了白色的部分
        total = cv2.countNonZero(mask)
        print('total=',total)
        #判断一下,如果他涂了两个选项,应该另外处理
        if total>1000:
            count += 1
        if count< 2:
            # 通过阈值判断,如果是第一次/当前值比上一次的大,则bubbled记录下来,直到最后bubbled就是一组中值最大的那个选项
            if bubbled is None or total > bubbled[0]:
                bubbled = (total, j)
        else:
            print('题目[%d]选项大于1'.format(j))
    print('-------------------')
    count=0


    # 对比正确答案
    color = (0, 0, 255)
    k = ANSWER_KEY[q]

    # 判断正确,bubbled=(656,1),其中第二个依次表示A,B,C,D,E,
    if k == bubbled[1]:
        color = (0, 255, 0)
        correct += 1

    # 绘图
    cv2.drawContours(imgWrap, [cnts[k]], -1, color, 3)

score = (correct / 5.0) * 100
print("[INFO] score: {:.2f}%".format(score))
cv2.putText(imgWrap, "{:.2f}%".format(score), (10, 30),
            cv2.FONT_HERSHEY_SIMPLEX, 0.9, (0, 0, 255), 2)
cv2.imshow("Original", imgOriginal)
cv2.imshow("Exam", imgWrap)
cv2.waitKey()

踩坑记录

1是在进行透视变换时,src和dst必须为float32类型,否则会报错

2是在给src赋值的时候,不必一个个的赋值,可以直接使用np.array

src = np.array(wraped, dtype="float32")

3是获取图像的行和列不是C++中的rows和cols,而是

w = imgOriginal.shape[0]h = imgOriginal.shape[1]
原文地址:https://www.cnblogs.com/hongweijiang/p/14931426.html