最近在Github看到一个很有趣的案例，基于opencv与KNN的数独自动识别与填充算法，想自己写一下跑一遍，先上原地址：
【转】
上效果图：
Processing:
- 1:九宫格数字提取---Extract_nums.py - **方法**: cv2.findContours(image, mode, method[, contours[, hierarchy[, offset] ] ]) → contours, hierarchy ~ Python中, findContours()接收如下参数并返回contours和hierarchy(等级) ~~1.image 源图像，一般为8为单通道图像，更具体来说，二值图像。其他情况暂且不论。 ~~2.mode 轮廓检索模式，简要介绍几种： (1)cv2.RETR_EXTERNAL 只检测外轮廓。对所有轮廓设置hierarchy[i][2]=hierarchy[i][3]=-1 (2)cv2.RETR_LIST 提取所有轮廓，并放置在list中，检测到的轮廓不建立等级关系。 (3)cv2.RETR_TREE 提取所有轮廓，建立网状的轮廓结构。 ~~3.method 轮廓的近似办法，是提取轮廓上所有像素点，还是只提取关键的一些点。比如一条线段是提取所有点还是只提取两个端点。 ~~4.contours 检测到的轮廓，为组成轮廓的点集。 ~~5.hierarchy (1)meanings: 什么是层级结构？ 检测轮廓时，一个轮廓包含了另一个轮廓（同心圆）>>>外圆：父轮廓；内圆：子轮廓。 同一个级别又分为前一个轮廓，后一个轮廓。综上所述，hierachy表达不同轮廓之间的关系。 (2)对每一个轮廓有：[Next,Previous,First_Child,Parent]) (3)cv2.RETR_EXTERNAL 只检测外轮廓。对所有轮廓设置hierarchy[i][2]=hierarchy[i][3]=-1，即(2)中First_Child和Parent 都是-11 # -*- coding: UTF-8 -*- 2 import cv2 3 import glob as gb 4 # 九宫格，最外面包围所有的是0号轮廓，里面81个小方格就是0号轮廓的自轮廓， 5 # 而每一个已知数字的轮廓都是对应方格的子轮廓。 6 7 # 对每一个轮廓有：[Next,Previous,First_Child,Parent]) 8 9 #read the pic10 img = cv2.imread('images/001.jpg')11 #transfer the BGR to GRAY12 # gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)13 gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)14 15 #阈值分割16 ret, thresh = cv2.threshold(gray,200,255,1)17 18 #对二值图像进行膨胀操作???19 kernel = cv2.getStructuringElement(cv2.MORPH_CROSS,(5,5))20 dilated = cv2.dilate(thresh,kernel)21 22 #轮廓提取，cv2.RETR_TREE表示建立层级结构23 image, contours, hierarchy = cv2.findContours(dilated,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)24 print(hierarchy.shape)25 # cv2.namedWindow("img", cv2.WINDOW_NORMAL); 26 # cv2.imshow("img", image)27 # cv2.waitKey(0) 28 # 提取小方格,父轮廓都为0号29 boxes = []30 for i in range(len(hierarchy[0])):31 if hierarchy[0][i][3] == 0:32 boxes.append(hierarchy[0][i])33 34 # #提取数字35 number_boxes = []36 for j in range(len(boxes)):37 if boxes[j][2] != -1:38 #number_boxes.append(boxes[j])39 x,y,w,h = cv2.boundingRect(contours[boxes[j][2]])40 number_boxes.append([x,y,w,h])41 img = cv2.rectangle(img,(x-1,y-1),(x+w+1,y+h+1),(0,0,255),2)42 43 cv2.namedWindow("img", cv2.WINDOW_NORMAL); 44 cv2.imshow("img", img)45 cv2.waitKey(0) - 2:数字识别
- a.数据收集和处理---build_datasets.py - b.kNN数字识别---knn_recogntion.py -(1) 这里用opencv自带的knn算法实现。我同时尝试了opencv自带的神经网络和SVM，发现还是kNN的效果最好。有兴趣的可以自己去尝试一下。也可能是我参数没调好。 -(2) -1.加载上面保存的样本和标签数据； -2.分别用80个作为训练数据，20个作为测试数据； -3.用opencv自带的knn训练模型； -4.用训练好的模型识别测试数据中的数字； -5.输出预测值和实际标签值。1 # -*- coding: UTF-8 -*- 2 import numpy as np 3 import cv2 4 import glob as gb 5 6 #获取文件夹下原始数字图片 7 img_path = gb.glob("numbers\\*") 8 # img_path = cv2.imread() 9 10 11 k = 0 12 labels = [] 13 samples = [] 14 15 #1.遍历文件夹下原始数字图片 16 #2.对每张图片进行轮廓提取操作，只提取外围轮廓 17 #3.求轮廓外包矩形，并根据矩形大小信息筛选出所有的数字轮廓 18 #4.然后根据位置信息对数字框排序，显然第一排依次是12345，第二排依次是67890； 19 #5.提取每一个数字所在的矩形框，作为ROI取出 20 # 对每一个轮廓等级有：[Next,Previous,First_Child,Parent]) 21 22 # for path in range(10): 23 for path in img_path: 24 # print(path) 25 # img = cv2.imread('numbers/'+str(path+1)+'.jpg') 26 img = cv2.imread(path) 27 #img's shape: (pixel,pixel, 3 channels--> B,G,R) 28 # print("image's shape:{}"+str(img.shape)) 29 #转化为灰度图像 30 gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY) 31 32 # #高斯滤波 33 blur = cv2.GaussianBlur(gray,(5,5),0) 34 35 #adaptive Threshold自适应阈值 36 #将灰度图像转换为二值图像 37 thresh = cv2.adaptiveThreshold(blur,255,1,1,11,2) 38 39 #提取轮廓 40 image,contours,hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE) 41 print("hierarchy's shape:{}".format(hierarchy.shape)) 42 # print("contours's shape:{}".format(contours.shape)) 43 # print(, sep, end, file, flush) 44 45 #求轮廓外包矩形，并根据矩形大小信息筛选出所有的数字轮廓 46 height,width = img.shape[:2] 47 w = width/5 48 recl_list = [] 49 list1 = [] 50 list2 = [] 51 # j = 0 52 # labels =[] 53 for cnt in contours: 54 #计算并返回指定点集的最小边界矩形 55 [x,y,w,h] = cv2.boundingRect(cnt) 56 57 if w>30 and h>(height/4): 58 if y<(height/2): 59 list1.append([x,y,w,h]) 60 else: 61 list2.append([x,y,w,h]) 62 63 # key = lambda 元素: 元素[字段索引] x:x[]可以随意修改，排序方式按照括号[]里面的维度进行 64 # [0] 第一维排序、 [1]第一维排序...以此类推 65 list1_sorted = sorted(list1, key=lambda t:t[0]) 66 list2_sorted = sorted(list2, key=lambda t:t[0]) 67 68 #提取出每一个数字所在的矩形框，作为ROI取出 69 for i in range(5): 70 [x1,y1,w1,h1] = list1_sorted[i] 71 [x2,y2,w2,h2] = list2_sorted[i] 72 number_roi1 = gray[y1:y1+h1, x1:x1+w1] #Cut the frame to size 73 number_roi2 = gray[y2:y2+h2, x2:x2+w2] #... 74 75 #数据预处理 76 #1. 把每一张ROI转换为40x20 ##########注意Opencv 里是宽x长 77 resized_roi1 = cv2.resize(number_roi1,(20,40)) 78 resized_roi2 = cv2.resize(number_roi2,(20,40)) 79 80 #2.阈值分割 灰度图-> 0-1二值图 81 thresh1 = cv2.adaptiveThreshold(resized_roi1,255,1,1,11,2) 82 thresh2 = cv2.adaptiveThreshold(resized_roi2,255,1,1,11,2) 83 84 #3.创建文件夹 85 # path_1 = "F:\\opencv_learning\\opencv_knn_soduko\\New_datasets\\"+str(i+1)+"\\"+str(k)+".jpg" 86 # j = 0 87 number_path1 = "datasets_hzc\\%s\\%d" % (str(i+1),k) + '.jpg' 88 89 j = i+6 90 if j == 10: 91 j = 0 92 # path_2 = "F:\\opencv_learning\\opencv_knn_soduko\\New_datasets\\"+str(j)+"\\"+str(k)+".jpg" 93 number_path2 = "datasets_hzc\\%s\\%d" % (str(j),k) + '.jpg' 94 k+=1 95 96 #4.Nomalized 97 # normalized_roi1 98 normalized_roil = thresh1/255 99 normalized_roi2 = thresh2/255100 # 5.write into the files101 # P.S Don't forget to annotate them, because u dont have to operate it twice102 # cv2.imwrite(number_path1,thresh1)103 # cv2.imwrite(number_path2,thresh2)104 #把处理完的二值图像展开成一行，以便knn处理105 sample_1 = normalized_roil.reshape((1,800))106 samples.append(sample_1[0])107 labels.append(float(i+1))108 109 #保存sample供训练用110 sample_2 = normalized_roi2.reshape((1,800))111 samples.append(sample_2[0])112 #把数字标签按照数字的保存顺序对应保存成训练用的数据113 labels.append(float(j))114 115 cv2.imwrite(number_path1,thresh1)116 cv2.imwrite(number_path2,thresh2)117 cv2.imshow("number",normalized_roil)118 cv2.waitKey(5)119 # cv2.imshow("1", thresh1)120 # cv2.imshow("2", thresh2)121 # cv2.waitKey(300)122 # print(list1_sorted)123 # print(list2_sorted)124 # cv2.imshow("train_pic", image)125 # cv2.waitKey(300)126 127 #保运数据供KNN用 k邻近算法128 print(np.array(labels).shape)129 130 print("\n"*3)131 samples = np.array(samples,np.float32)132 #(100,800) 1x100, 800133 print("train_data's dimension:{}".format(samples.shape))134 labels = np.array(labels,np.float32)135 labels = labels.reshape((labels.size,1))136 print("labels's dimension:{}".format(labels.shape))137 np.save('samples,npy',samples)138 np.save('label.npy',labels) - 3.数独生成和求解
1 # -*- coding: UTF-8 -*- 2 import numpy as np 3 import cv2 4 5 # 利用K邻近算法来进行数字识别 6 # 1.加载上面保存的样本和标签数据； 7 # 2.分别用80个作为训练数据，20个作为测试数据； 8 # 3.用opencv自带的knn训练模型； 9 # 4.用训练好的模型识别测试数据中的数字； 10 # 5.输出预测值和实际标签值。 11 12 # print(labels.T) 13 # print(np.sort(labels, axis=0, kind='quicksort') 14 # print("samples's shape:{}".format(samples.shape)) 15 16 ## 数独求解算法，回溯法。来源见下面链接，有细微改动。 17 ## http://stackoverflow.com/questions/1697334/algorithm-for-solving-sudoku 18 def findNextCellToFill(grid, i, j): 19 for x in range(i,9): 20 for y in range(j,9): 21 if grid[x][y] == 0: 22 return x,y 23 for x in range(0,9): 24 for y in range(0,9): 25 if grid[x][y] == 0: 26 return x,y 27 return -1,-1 28 29 def isValid(grid, i, j, e): 30 rowOk = all([e != grid[i][x] for x in range(9)]) 31 if rowOk: 32 columnOk = all([e != grid[x][j] for x in range(9)]) 33 if columnOk: 34 # finding the top left x,y co-ordinates of the section containing the i,j cell 35 secTopX, secTopY = 3 *int(i/3), 3 *int(j/3) 36 for x in range(secTopX, secTopX+3): 37 for y in range(secTopY, secTopY+3): 38 if grid[x][y] == e: 39 return False 40 return True 41 return False 42 43 def solveSudoku(grid, i=0, j=0): 44 i,j = findNextCellToFill(grid, i, j) 45 if i == -1: 46 return True 47 for e in range(1,10): 48 if isValid(grid,i,j,e): 49 grid[i][j] = e 50 if solveSudoku(grid, i, j): 51 return True 52 # Undo the current cell for backtracking 53 grid[i][j] = 0 54 return False 55 56 samples = np.load('F:/opencv_learning/opencv-soduko-master/samples.npy') 57 labels = np.load('F:/opencv_learning/opencv-soduko-master/label.npy') 58 59 k=80 60 train_label = labels[:k] 61 test_label = labels[k:] 62 train_input = samples[:k] 63 test_input = samples[k:] 64 print("train_label's shape:{}".format(train_label.shape)) 65 print("train_label's shape:{}".format(train_input.shape)) 66 print("test_label's shape:{}".format(test_label.shape)) 67 print("test_input's shape:{}".format(test_input.shape)) 68 69 #create the model 70 model = cv2.ml.KNearest_create() 71 model.train(train_input,cv2.ml.ROW_SAMPLE,train_label) 72 73 #retval:返回值类型说明 74 # retval,results,neigh_resp,dists = model.findNearest(test_input,1) 75 # string = results.ravel() 76 # print(string) 77 # print(test_label.reshape(1,len(test_label))[0]) 78 # print(string.shape) 79 # print("The original label:{}".format(test_label.reshape(1,len(test_label))[0])) 80 # print("The prediction:{}".format(str(string))) 81 82 # # print(test_label.T) 83 # count = 0 84 # string = string.reshape(1,len(string)) 85 # print(string.shape) 86 # string = np.array(string).T 87 88 # for index,value in enumerate(test_label): 89 # if value != string[index]: 90 # count+=1 91 # print(count) 92 93 # accuracy = 1 - (count / len(string)) 94 # print("The test accuracy:{}".format(accuracy)) 95 96 img = cv2.imread('./images/001.jpg') 97 gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY) 98 99 ##阈值分割100 ret,thresh = cv2.threshold(gray,200,255,1)101 102 # in ord to do morphological operation: returns a structuring element103 kernel = cv2.getStructuringElement(cv2.MORPH_CROSS,(5,5))104 #Brief dilated an image by a specific structuring element105 #膨胀106 dilated = cv2.dilate(thresh,kernel)107 108 #提取轮廓109 image, contours, hierarchy = cv2.findContours(dilated,cv2.RETR_TREE,\110 cv2.CHAIN_APPROX_SIMPLE)111 112 #提取81个小方格113 #hierarchy[Next,Previous,First_Child,Parent])114 boxes = []115 for i in range(len(hierarchy[0])):116 if hierarchy[0][i][3] == 0:117 boxes.append(hierarchy[0][i])118 119 print("81个小方格: dimension:{}".format(np.array(boxes).shape))120 # print(boxes)121 122 height,width = img.shape[:2]123 124 #"/9" 9行9列125 box_h = height/9126 box_w = height/9127 number_boxes = []128 129 #数独转化为零阵130 soduko = np.zeros((9,9),np.int32)131 132 for j in range(len(boxes)):133 if boxes[j][2] != -1:134 #Calculates and returns the minimal up-right bounding 135 #rectangle for the specified point set136 x,y,w,h = cv2.boundingRect(contours[boxes[j][2]])137 number_boxes.append([x,y,w,h])138 #process the data that was extracted139 number_roi = gray[y:y+h,x:x+w]140 #unit the size 141 resized_roi = cv2.resize(number_roi,(20,40))142 thresh1 = cv2.adaptiveThreshold(resized_roi,255,1,1,11,2)143 #归一化像素值144 normalized_roi = thresh1/255145 146 #展开一行让knn识别147 sample1 = normalized_roi.reshape((1,800))148 sample1 = np.array(sample1,np.float32)149 150 retval,result,neigh_resp,dists = model.findNearest(sample1,1)151 number = int(result.ravel()[0])152 # print(results.ravel())153 154 #识别结果155 cv2.putText(img,str(number),(x+w+1,y+h-20), 3, 2., (255, 0, 0), 2, cv2.LINE_AA)156 157 #矩阵中位置158 soduko[int(y/box_h)][int(x/box_w)] = number159 160 cv2.namedWindow("img", cv2.WINDOW_NORMAL)161 cv2.imshow("img", img)162 # cv2.imshow("normalized_roi",normalized_roi)163 cv2.waitKey(120)164 165 print("\n生成的数独\n")166 print(soduko)167 print("\n求解后的数独\n")168 169 #数独求解170 solveSudoku(soduko)171 print(soduko)172 173 print("\n验算：求每行每列的和\n")174 #map求和 map根据指定的函数对指定序列做映射175 #map(function, iterable)176 # row_sum = map(sum,soduko)177 # col_sum = map(sum,zip(*soduko))178 print(soduko.shape)179 # np.sum(0>>row, 1/-1>>column)180 row_sum = np.sum(soduko,axis = 0)181 col_sum = np.sum(soduko,axis = -1,keepdims=1)182 print(row_sum)183 print(col_sum.T)184 185 #把对应结果按照位置填充图片中186 for i in range(9):187 for j in range(9):188 x = int((i+0.25)*box_w) 189 y = int((j+0.5)*box_h)190 #brief Draws a text string.191 cv2.putText(img,str(soduko[j][i]),(x,y), 3, 2.5, (0, 0, 255), 2, cv2.LINE_AA)192 193 cv2.namedWindow("img",cv2.WINDOW_NORMAL)194 cv2.imshow("img", img)195 cv2.waitKey(10000)196 197 # np.sum(a, axis=None, dtype=None, out=None, keepdims=np._NoValue, initial=np._NoValue)198 # print(list(row_sum))199 # print(list(col_sum))200 201 # cv2.imshow("origin",img)202 # cv2.waitKey(300)203 # cv2.imshow("gray",gray)204 # cv2.waitKey(300)205 # cv2.imshow("thresh",thresh)206 # cv2.imshow("dilated",dilated)207 # cv2.waitKey(0)208 # cv2.209 #SVM210 # C = 5211 # gamma = 0.5212 # model_svm = cv2.ml.SVM_create()213 # model_svm.setGamma(gamma)214 # model_svm.setC(C)215 216 # #LINEAR MODEL217 # model_svm.setKernel(cv2.ml.SVM_LINEAR)218 # model_svm.setType(cv2.ml.SVM_C_SVC)219 # model_svm.train(train_input, cv2.ml.ROW_SAMPLE, train_label)220 # predict_label = model_svm.predict(test_input)[1].ravel()221 # print(predict_label)222 # print(test_label.reshape(1,len(test_label))[0])