基于opencv的试卷检测识别（opencv，Tensorflow）

最近做了一个基于opencv的试卷识别项目，在此分享一下。

文末附有源码。

视频观看：

20211212

主界面：

选择图片后：

识别以后：

识别完成以后会自动截取不同的题目，然后保持到不同的文件夹中，分数会单独的保存到一个txt文本中。

手写数字数据集：

手写数字识别代码，建议不要直接用mnist手写数字数据集，因为使用这个数据集训练出来的网络，根本识别不了我自己写的数字，亲身体验！没办法，我只好自己制作了手写数据集，其实很简单。数据集如下所示：



其余的都差不多是这样，就不过多展示了。需要注意的是图片上面只有数字是黑色的，这样方面提取出数字，如果写错了，可以用一些图像编辑的软件将错的部分涂成白色即可，就和上图一样。
手写数字识别代码：

import tensorflow as tf
import cv2 as cv
import numpy as np
from get_data import *

model=tf.keras.models.Sequential([
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128,activation='relu'),
    tf.keras.layers.Dense(128,activation='relu'),
    tf.keras.layers.Dense(128,activation='relu'),
    tf.keras.layers.Dense(10,activation='softmax')
])
model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False),
              metrics=['sparse_categorical_accuracy'])
model.fit(x_train,y_train,batch_size=32,epochs=10,validation_split=0.1,validation_freq=1)
model.save('mode_2.h5')

是不是很简单呢？请看第四行：

from get_data import *

关于图像处理的部分在get_data.py这个代码中，下面是get_data.py的代码。

get_data.py

import numpy as np
import cv2 as cv

x_train=[]
y_train=[]


aa=9
for aa in range(10):
    src=cv.imread('data/{}.jpg'.format(aa))
    gray=cv.cvtColor(src,cv.COLOR_BGR2GRAY)
    thred=np.where(gray>150,255,0).astype('uint8')
    thred=255-thred
    #开闭运算
    k = np.ones((3, 3), np.uint8)
    #thred=cv.dilate(thred,k)
    thred = cv.morphologyEx(thred, cv.MORPH_CLOSE, k)

    cnts=cv.findContours(thred,cv.RETR_EXTERNAL,cv.CHAIN_APPROX_SIMPLE)[0]
    print(len(cnts))

    for i in cnts:
        area = cv.contourArea(i)
        if aa!=1:
            b=20
        else:b=15
        if area>=b:
            x, y, w, h = cv.boundingRect(i)
            cv.rectangle(src,(x,y),(x+w,y+h),(0,0,255),2)
            lkuo=thred[y:y+h,x:x+w]
            da = max(h, w)
            rate = da / 40
            ro = cv.resize(lkuo, (int(w / rate), int(h / rate)))
            h, w = ro.shape
            t, b = int((43 - h) / 2), 43 - h - int((43 - h) / 2)
            l, r = int((43 - w) / 2), 43 - w - int((43 - w) / 2)

            ro = cv.copyMakeBorder(ro, t, b, l, r, cv.BORDER_CONSTANT, value=0)

            ro = cv.resize(ro, (40, 40))
            ro = np.where(ro > 0, 255, 0).astype('float32')
            ro = ro / 255
            x_train.append(ro)
            y_train.append(aa)

x_train=np.array(x_train).astype('float32')
y_train=np.array(y_train).astype('float32')
y_train=np.reshape(y_train,(y_train.shape[0],1))
np.random.seed(1)
np.random.shuffle(x_train)
np.random.seed(1)
np.random.shuffle(y_train)
print(y_train)
if __name__=='__main__':
    pass

好了，到这里手写数字识别的部分已经完结了，后面可以直接调用我们训练好的网络就可以识别了，下面是关于试卷检测的代码：

import time

import cv2 as cv
import numpy as np
from tensorflow.keras.models import load_model
import os

model =load_model('mode.h5')
def zb(img,a):
    # b=np.sort(a,axis=0)
    idx = np.argsort(a, axis=0)
    aa = a[idx[:, 0]]
    idx12=np.argsort(aa[:2],axis=0)
    idx34 = np.argsort(aa[2:], axis=0)
    aa[:2]=aa[:2][idx12[:,1]]
    aa[2:]=aa[2:][idx34[:,1]]
    p1 = aa[0]
    p2 = aa[1]
    p3 = aa[3]
    p4 = aa[2]
    # rect=[p1,p2,p3,p4]
    # rect=np.array(rect)
    w=max(np.sqrt(np.sum(np.square(p4-p1))),np.sqrt(np.sum(np.square(p3-p2))))
    h=max(np.sqrt(np.sum(np.square(p2-p1))),np.sqrt(np.sum(np.square(p3-p4))))
    dst=np.array([[0,0],
                  [w - 1, 0],
                  [w - 1, h - 1],
                  [0, h - 1]],dtype='float32')

    xx=[p1,p4,p3,p2]
    aa=np.array(xx).astype('float32')
    M=cv.getPerspectiveTransform(aa,dst)
    warped=cv.warpPerspective(img,M,(int(w),int(h)))
    return warped
def draw(img,x1,y1,x2,y2,text=None,dr=True):
    if dr:
        cv.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 1)
        cv.putText(img,text,(x1,y1-15),cv.FONT_HERSHEY_SIMPLEX,1,(0,255,0),2)
    src=warped_copy[y1:y2,x1:x2]
    return src
def chang_pic(ro):
    h,w=ro.shape
    da=max(h,w)
    rate=da/40
    ro=cv.resize(ro,(int(w/rate),int(h/rate)))
    h,w=ro.shape
    t,b=int((43-h)/2),43-h-int((43-h)/2)
    l,r=int((43-w)/2),43-w-int((43-w)/2)

    ro=cv.copyMakeBorder(ro,t,b,l,r,cv.BORDER_CONSTANT,value=0)

    ro=cv.resize(ro,(40,40))
    #cv.imshow('ro1', ro)
    ro=np.where(ro>0,255,0).astype('float32')
    ro=ro/255

    print('ro=',ro.shape)
    ro=np.reshape(ro,(1,40,40))
    pre=model.predict(ro)[0]
    true=np.argmax(pre)
    return str(true)

def shuzi(imgs):
    imgs=imgs[4:-4,4:-4]
    cc=max(imgs.shape[0],imgs.shape[1])
    imgs=cv.resize(imgs,(cc,cc))
    imgs=cv.cvtColor(imgs,cv.COLOR_BGR2GRAY)
    #cv.imshow('gray',imgs)
    thred = np.where(imgs > 215, 0, 255).astype('uint8')
    #cv.imshow('a11',thred)
    contours, hierarchy = cv.findContours(thred, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_NONE)
    lis=[]
    if len(contours)==0:
        return None
    contours_2=[]
    for i in contours:
        x, y, w, h = cv.boundingRect(i)
        area = cv.contourArea(i)
        if area>10:
            lis.append(x)
            contours_2.append(i)
        print(area)
    lis=np.array(lis)
    #contours_2=np.array(contours_2)
    idx=np.argsort(lis)
    print(idx)
    #contours_2=np.array(contours_2)
    contours_3=[]
    for id,ii in enumerate(idx):
        contours_3.append(contours_2[ii])
    #print(contours_2)
    ll=''
    for j in contours_3:
        x1, y1, w1, h1 = cv.boundingRect(j)
        lunkuo=thred[y1:y1+h1,x1:x1+w1]
        #cv.imshow('lh45',lunkuo)
        number=chang_pic(lunkuo)
        ll=ll+number
    return ll

nam='3.jpg'
image=cv.imread('./pic/'+nam)
def mmain(image):
    a=1000
    ratio=image.shape[0]/a
    orig=image.copy()

    image=cv.resize(image,(int(image.shape[1]/ratio),a))

    gray=cv.cvtColor(image,cv.COLOR_BGR2GRAY)
    gray=cv.GaussianBlur(gray,(5,5),0)
    #边缘检测
    edged=cv.Canny(gray,75,200)
    cv.imwrite('lk.jpg',edged)

    #轮廓检测
    cnts=cv.findContours(edged.copy(),cv.RETR_LIST,cv.CHAIN_APPROX_SIMPLE)[0]
    cnts=sorted(cnts,key=cv.contourArea,reverse=True)[0]
    #遍历轮廓
    # cnts=np.reshape(cnts,(cnts.shape[0],cnts.shape[2]))
    # print(cnts.shape)
    peri=cv.arcLength(cnts,True)
    approx=cv.approxPolyDP(cnts,0.02*peri,True)


    if len(approx)==4:
        screenCnt=approx
        cv.drawContours(image,[screenCnt],-1,(0,255,0),2)
        warped=zb(orig,screenCnt.reshape(4,2)*ratio)

        warped=cv.resize(warped,(724,1000))
        global warped_copy
        warped_copy=warped.copy()
        #cv.imwrite('pic/sjuan.jpg',warped)
        print(warped.shape)

        score1=draw(warped,96,85,138,125)
        score2=draw(warped,150,85,193,125)
        score3=draw(warped,200,85,245,125)
        score4=draw(warped,254,85,296,125)
        #识别分数
        num1 = shuzi(score1)

        print('num1=',num1)
        num2 = shuzi(score2)
        print(num2)
        num3 = shuzi(score3)
        print('num3=',num3)
        num4 = shuzi(score4)
        print('num4=',num4)
        s1= draw(warped, 96, 85, 138, 125,text=num1)
        s2 = draw(warped, 150, 85, 193, 125,text=num2)
        s3 = draw(warped, 200, 85, 245, 125,text=num3)
        s4 = draw(warped, 254, 85, 296, 125,text=num4)

        dati1_1=draw(warped,25,293,335,448,text='one')
        dati1_2=draw(warped,350,42,660,110,text='one')
        dati1=np.concatenate((dati1_1,dati1_2),axis=0)
        dati2=draw(warped,350,110,660,280,text='two')
        dati3=draw(warped,350,283,660,450,text='three')
        dati4=draw(warped,28,541,416,916,text='four')

        da1_1=draw(warped_copy,37,294,323,336,dr=False)
        da1_2 = draw(warped_copy, 37, 336, 323, 368, dr=False)
        da1_3 = draw(warped_copy, 37, 368, 323, 390, dr=False)
        da1_4= draw(warped_copy, 37, 390, 323, 442, dr=False)
        da1_5 = draw(warped_copy, 353, 46, 651, 111, dr=False)
        ddaa1=[da1_1,da1_2,da1_3,da1_4,da1_5]
        ddaa1_name=['da1_1','da1_2','da1_3','da1_4','da1_5']

        da2_1 = draw(warped_copy, 353, 120, 651, 166, dr=False)
        da2_2 = draw(warped_copy, 353, 166, 651, 192, dr=False)
        da2_3 = draw(warped_copy, 353, 195, 651, 218, dr=False)
        da2_4 = draw(warped_copy, 353, 222, 651, 252, dr=False)
        da2_5 = draw(warped_copy, 353, 250, 651, 282, dr=False)
        ddaa2=[da2_1,da2_2,da2_3,da2_4,da2_5]
        ddaa2_name=['da2_1','da2_2','da2_3','da2_4','da2_5']

        da3_1 = draw(warped_copy, 353, 281, 634, 322, dr=False)
        da3_2 = draw(warped_copy, 353, 322, 634, 346, dr=False)
        da3_3 = draw(warped_copy, 353, 346, 634, 378, dr=False)
        da3_4 = draw(warped_copy, 353, 378, 634, 400, dr=False)
        da3_5 = draw(warped_copy, 353, 400, 634, 442, dr=False)
        ddaa3=[da3_1,da3_2,da3_3,da3_4,da3_5]
        ddaa3_name=['da3_1','da3_2','da3_3','da3_4','da3_5']


        da4_1 = draw(warped_copy, 35, 551, 381, 617, dr=False)
        da4_2 = draw(warped_copy, 35, 617, 381, 700, dr=False)
        da4_3 = draw(warped_copy, 35, 700, 381, 786, dr=False)
        da4_4 = draw(warped_copy, 35, 786, 381, 912, dr=False)
        ddaa4=[da4_1,da4_2,da4_3,da4_4]
        ddaa4_name=['da4_1','da4_2','da4_3','da4_4']



        t1_1=draw(warped,97,136,140,160)
        t1_2=draw(warped,97,168,140,190)
        t1_3 = draw(warped, 97, 200, 140, 224)
        t1_4 = draw(warped, 97, 232, 140, 255)
        t1_5 = draw(warped, 97, 262, 140, 287)


        t2_1 = draw(warped, 150, 136, 192, 160)
        t2_2 = draw(warped, 150, 168, 192, 190)
        t2_3 = draw(warped, 150, 200, 192, 224)
        t2_4 = draw(warped, 150, 232, 192, 255)
        t2_5 = draw(warped, 150, 262, 192, 287)

        t3_1 = draw(warped, 200, 136, 245, 160)
        t3_2 = draw(warped, 200, 168, 245, 190)
        t3_3 = draw(warped, 200, 200, 245, 224)
        t3_4 = draw(warped, 200, 232, 245, 255)
        t3_5 = draw(warped, 200, 262, 245, 287)

        t4_1 = draw(warped, 253, 136, 297, 160)
        t4_2 = draw(warped, 253, 168, 297, 190)
        t4_3 = draw(warped, 253, 200, 297, 224)
        t4_4 = draw(warped, 253, 232, 297, 255)
        t4_5 = draw(warped, 253, 262, 297, 287)

        fen=[t1_1,t1_2,t1_3,t1_4,t1_5,
             t2_1,t2_2,t2_3,t2_4,t2_5,
             t3_1,t3_2,t3_3,t3_4,t3_5,
             t4_1,t4_2,t4_3,t4_4,t4_5]
        name=['t1_1','t1_2','t1_3','t1_4','t1_5',
             't2_1','t2_2','t2_3','t2_4','t2_5',
             't3_1','t3_2','t3_3','t3_4','t3_5',
             't4_1','t4_2','t4_3','t4_4','t4_5']

        ss=[num1,num2,num3,num4]
        fensu=0
        for iii in ss:
            try:
                fensu+=int(iii)
            except:pass
        print(fensu)
        #创建文件夹保存文件
        if not os.path.isdir("timu"):
            os.mkdir("timu")
        if not os.path.isdir("timu/score"):
            os.mkdir("timu/score")
        if not os.path.isdir("timu/one"):
            os.mkdir("timu/one")
        if not os.path.isdir("timu/two"):
            os.mkdir("timu/two")
        if not os.path.isdir("timu/three"):
            os.mkdir("timu/three")
        if not os.path.isdir("timu/four"):
            os.mkdir("timu/four")
        na=nam.split('.')[0]

        #保存大题
        for sce,wq in zip(ddaa1,ddaa1_name):
            cv.imwrite('timu/one/{}_{}.jpg'.format(na,wq),sce)
        for sce,wq in zip(ddaa2,ddaa2_name):
            cv.imwrite('timu/two/{}_{}.jpg'.format(na,wq),sce)
        for sce,wq in zip(ddaa3,ddaa3_name):
            cv.imwrite('timu/three/{}_{}.jpg'.format(na,wq),sce)
        for sce,wq in zip(ddaa4,ddaa4_name):
            cv.imwrite('timu/four/{}_{}.jpg'.format(na,wq),sce)


        #cv.imwrite('timu/{}_one.jpg'.format(na),dati1)
        # cv.imwrite('timu/{}_two.jpg'.format(na), dati2)
        # cv.imwrite('timu/{}_three.jpg'.format(na), dati3)
        # cv.imwrite('timu/{}_four.jpg'.format(na), dati4)

        #保存分数区域
        for ax,nna in zip(fen,name):
            cv.imwrite('timu/score/{}_{}.jpg'.format(na,nna), ax)
        #保存题目

        # cv.imwrite('warp.jpg',warped_copy)
        with open('score.txt','a') as f:
            f.write(str(num1)+'  '+str(num2)+'  '+str(num3)+'  '+str(num4)+'  '+str(fensu)+'\n')
        return warped,image
    else:
        return
if __name__=='__main__':
    img=cv.imread('./pic/2.jpg')
    warped, img = mmain(img)
    cv.imshow('q12',img)
    cv.waitKey(0)
    cv.destroyWindow()

下面是界面的部分代码：

import sys, cv2
from PyQt5.QtGui import *
from PyQt5.QtWidgets import *
from PyQt5.QtCore import *
from untitled import Ui_Dialog
from t1 import *

class My(QMainWindow,Ui_Dialog):
    def __init__(self):
        super(My,self).__init__()
        self.setupUi(self)
        self.pushButton.clicked.connect(self.pic)

        self.setWindowTitle('试卷检测')
        self.pushButton_2.clicked.connect(self.dis)
        self.setIcon()
    def setIcon(self):
        palette1 = QPalette()
        # palette1.setColor(self.backgroundRole(), QColor(192,253,123))   # 设置背景颜色
        palette1.setBrush(self.backgroundRole(), QBrush(QPixmap('22.png')))  # 设置背景图片
        self.setPalette(palette1)
        # self.setAutoFillBackground(True) # 不设置也可以

        # self.setGeometry(300, 300, 250, 150)
        #self.setWindowIcon(QIcon('22.jpg'))

    def pic(self):
        imgName, imgType = QFileDialog.getOpenFileName(self,
                                                       "打开图片",
                                                       "",
                                                       " *.jpg;;*.png;;*.jpeg;;*.bmp;;All Files (*)")

        img = cv2.imread(imgName)

        self.warped,self.img=mmain(img)
        h1,w1=self.warped.shape[0],self.warped.shape[1]
        self.warped=cv.resize(self.warped,(int(w1/(h1/750)),750))
        print(self.warped.shape)
        self.img = cv.resize(self.img, (int(w1 / (h1 / 750)), 750))
        try:
            self.warped=self.cv_qt(self.warped)
            self.img = self.cv_qt(self.img)
            self.label.setPixmap(QPixmap.fromImage(self.img))
        except:pass

    def cv_qt(self, src):
        h, w, d = src.shape
        bytesperline = d * w
        # self.src=cv.cvtColor(self.src,cv.COLOR_BGR2RGB)
        qt_image = QImage(src.data, w, h, bytesperline, QImage.Format_RGB888).rgbSwapped()
        return qt_image
    def dis(self):
        self.label.setPixmap(QPixmap.fromImage(self.warped))

if __name__ == '__main__':
    app = QApplication(sys.argv)
    # 初始化GUI窗口 并传入摄像头句柄
    win = My()
    win.show()
    sys.exit(app.exec_())

整体目录结构：

运行main.py即可使用。