【深度學習系列】用PaddlePaddle進行車牌識別（一）
作者：Charlotte77
小伙伴們，終于到了實戰部分了！今天給大家帶來的項目是用PaddlePaddle進行車牌識別。車牌識別其實屬于比較常見的圖像識別的項目了，目前也屬于比較成熟的應用，大多數老牌廠家能做到準確率99%+。
小伙伴們，終于到了實戰部分了！今天給大家帶來的項目是用PaddlePaddle進行車牌識別。車牌識別其實屬于比較常見的圖像識別的項目了，目前也屬于比較成熟的應用，大多數老牌廠家能做到準確率99%+。傳統的方法需要對圖像進行多次預處理再用機器學習的分類算法進行分類識別，然而深度學習發展起來以后，我們可以通過用CNN來進行端對端的車牌識別。任何模型的訓練都離不開數據，在車牌識別中，除了晚上能下載到的一些包含車牌的數據是不夠的，本篇文章的主要目的是教大家如何批量生成車牌。
生成車牌數據
1.定義車牌數據所需字符
車牌中包括省份簡稱、大寫英文字母和數字，我們首先定義需要的字符和字典，方便后面使用
1 index = {"京": 0, "滬": 1, "津": 2, "渝": 3, "冀": 4, "晉": 5, "蒙": 6, "遼": 7, "吉": 8, "黑": 9, "蘇": 10, "浙": 11, "皖": 12,
2 "閩": 13, "贛": 14, "魯": 15, "豫": 16, "鄂": 17, "湘": 18, "粵": 19, "桂": 20, "瓊": 21, "川": 22, "貴": 23, "云": 24,
3 "藏": 25, "陜": 26, "甘": 27, "青": 28, "寧": 29, "新": 30, "0": 31, "1": 32, "2": 33, "3": 34, "4": 35, "5": 36,
4 "6": 37, "7": 38, "8": 39, "9": 40, "A": 41, "B": 42, "C": 43, "D": 44, "E": 45, "F": 46, "G": 47, "H": 48,
5 "J": 49, "K": 50, "L": 51, "M": 52, "N": 53, "P": 54, "Q": 55, "R": 56, "S": 57, "T": 58, "U": 59, "V": 60,
6 "W": 61, "X": 62, "Y": 63, "Z": 64};
7
8 chars = ["京", "滬", "津", "渝", "冀", "晉", "蒙", "遼", "吉", "黑", "蘇", "浙", "皖", "閩", "贛", "魯", "豫", "鄂", "湘", "粵", "桂",
9 "瓊", "川", "貴", "云", "藏", "陜", "甘", "青", "寧", "新", "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "A",
10 "B", "C", "D", "E", "F", "G", "H", "J", "K", "L", "M", "N", "P", "Q", "R", "S", "T", "U", "V", "W", "X",
11 "Y", "Z"
12 ];
2.生成中英文字符
1 def GenCh(f,val):
2 """
3 生成中文字符
4 """
5 img=Image.new("RGB", (45,70),(255,255,255))
6 draw = ImageDraw.Draw(img)
7 draw.text((0, 3),val,(0,0,0),font=f)
8 img = img.resize((23,70))
9 A = np.array(img)
10 return A
11
12 def GenCh1(f,val):
13 """
14 生成英文字符
15 """
16 img=Image.new("RGB", (23,70),(255,255,255))
17 draw = ImageDraw.Draw(img)
18 draw.text((0, 2),val.decode('utf-8'),(0,0,0),font=f)
19 A = np.array(img)
20 return A
3.對數據添加各種噪音和畸變，模糊處理
1 def AddSmudginess(img, Smu): 2 rows = r(Smu.shape[0] - 50) 3 cols = r(Smu.shape[1] - 50) 4 adder = Smu[rows:rows + 50, cols:cols + 50]; 5 adder = cv2.resize(adder, (50, 50)); 6 #adder = cv2.bitwise_not(adder) 7 img = cv2.resize(img,(50,50)) 8 img = cv2.bitwise_not(img) 9 img = cv2.bitwise_and(adder, img) 10 img = cv2.bitwise_not(img) 11 return img 12 13 14 def rot(img,angel,shape,max_angel): 15 """ 16 添加放射畸變 17 img 輸入圖像 18 factor 畸變的參數 19 size 為圖片的目標尺寸 20 """ 21 size_o = [shape[1],shape[0]] 22 size = (shape[1]+ int(shape[0]*cos((float(max_angel )/180) * 3.14)),shape[0]) 23 interval = abs( int( sin((float(angel) /180) * 3.14)* shape[0])); 24 pts1 = np.float32([[0,0],[0,size_o[1]],[size_o[0],0],[size_o[0],size_o[1]]]) 25 if(angel>0): 26 pts2 = np.float32([[interval,0],[0,size[1] ],[size[0],0 ],[size[0]-interval,size_o[1]]]) 27 else: 28 pts2 = np.float32([[0,0],[interval,size[1] ],[size[0]-interval,0 ],[size[0],size_o[1]]]) 29 M = cv2.getPerspectiveTransform(pts1,pts2); 30 dst = cv2.warpPerspective(img,M,size); 31 return dst 32 33 34 def rotRandrom(img, factor, size): 35 """ 36 添加透視畸變 37 """ 38 shape = size; 39 pts1 = np.float32([[0, 0], [0, shape[0]], [shape[1], 0], [shape[1], shape[0]]]) 40 pts2 = np.float32([[r(factor), r(factor)], [ r(factor), shape[0] - r(factor)], [shape[1] - r(factor), r(factor)], 41 [shape[1] - r(factor), shape[0] - r(factor)]]) 42 M = cv2.getPerspectiveTransform(pts1, pts2); 43 dst = cv2.warpPerspective(img, M, size); 44 return dst 45 46 def tfactor(img): 47 """ 48 添加飽和度光照的噪聲 49 """ 50 hsv = cv2.cvtColor(img,cv2.COLOR_BGR2HSV); 51 hsv[:,:,0] = hsv[:,:,0]*(0.8+ np.random.random()*0.2); 52 hsv[:,:,1] = hsv[:,:,1]*(0.3+ np.random.random()*0.7); 53 hsv[:,:,2] = hsv[:,:,2]*(0.2+ np.random.random()*0.8); 54 55 img = cv2.cvtColor(hsv,cv2.COLOR_HSV2BGR); 56 return img 57 58 def random_envirment(img,data_set): 59 """ 60 添加自然環境的噪聲 61 """ 62 index=r(len(data_set)) 63 env = cv2.imread(data_set[index]) 64 env = cv2.resize(env,(img.shape[1],img.shape[0])) 65 bak = (img==0); 66 bak = bak.astype(np.uint8)*255; 67 inv = cv2.bitwise_and(bak,env) 68 img = cv2.bitwise_or(inv,img) 69 return img 70 71 def AddGauss(img, level): 72 """ 73 添加高斯模糊 74 """ 75 return cv2.blur(img, (level * 2 + 1, level * 2 + 1)); 76 77 def r(val): 78 return int(np.random.random() * val) 79 80 def AddNoiseSingleChannel(single): 81 """ 82 添加高斯噪聲 83 """ 84 diff = 255-single.max(); 85 noise = np.random.normal(0,1+r(6),single.shape); 86 noise = (noise - noise.min())/(noise.max()-noise.min()) 87 noise= diff*noise; 88 noise= noise.astype(np.uint8) 89 dst = single + noise 90 return dst 91 92 def addNoise(img,sdev = 0.5,avg=10): 93 img[:,:,0] = AddNoiseSingleChannel(img[:,:,0]); 94 img[:,:,1] = AddNoiseSingleChannel(img[:,:,1]); 95 img[:,:,2] = AddNoiseSingleChannel(img[:,:,2]); 96 return img
4.加入背景圖片，生成車牌字符串list和label，并存為圖片格式，批量生成。
1 class GenPlate:
2
3 def __init__(self,fontCh,fontEng,NoPlates):
4 self.fontC = ImageFont.truetype(fontCh,43,0);
5 self.fontE = ImageFont.truetype(fontEng,60,0);
6 self.img=np.array(Image.new("RGB", (226,70),(255,255,255)))
7 self.bg = cv2.resize(cv2.imread("./images/template.bmp"),(226,70));
8 self.smu = cv2.imread("./images/smu2.jpg");
9 self.noplates_path = [];
10 for parent,parent_folder,filenames in os.walk(NoPlates):
11 for filename in filenames:
12 path = parent+"/"+filename;
13 self.noplates_path.append(path);
14
15
16 def draw(self,val):
17 offset= 2 ;
18 self.img[0:70,offset+8:offset+8+23]= GenCh(self.fontC,val[0]);
19 self.img[0:70,offset+8+23+6:offset+8+23+6+23]= GenCh1(self.fontE,val[1]);
20 for i in range(5):
21 base = offset+8+23+6+23+17 +i*23 + i*6 ;
22 self.img[0:70, base : base+23]= GenCh1(self.fontE,val[i+2]);
23 return self.img
24
25 def generate(self,text):
26 if len(text) == 9:
27 fg = self.draw(text.decode(encoding="utf-8"));
28 fg = cv2.bitwise_not(fg);
29 com = cv2.bitwise_or(fg,self.bg);
30 com = rot(com,r(60)-30,com.shape,30);
31 com = rotRandrom(com,10,(com.shape[1],com.shape[0]));
32 com = tfactor(com)
33 com = random_envirment(com,self.noplates_path);
34 com = AddGauss(com, 1+r(4));
35 com = addNoise(com);
36 return com
37
38 def genPlateString(self,pos,val):
39 '''
40 生成車牌String,存為圖片
41 生成車牌list,存為label
42 '''
43 plateStr = "";
44 plateList=[]
45 box = [0,0,0,0,0,0,0];
46 if(pos!=-1):
47 box[pos]=1;
48 for unit,cpos in zip(box,range(len(box))):
49 if unit == 1:
50 plateStr += val
51 #print plateStr
52 plateList.append(val)
53 else:
54 if cpos == 0:
55 plateStr += chars[r(31)]
56 plateList.append(plateStr)
57 elif cpos == 1:
58 plateStr += chars[41+r(24)]
59 plateList.append(plateStr)
60 else:
61 plateStr += chars[31 + r(34)]
62 plateList.append(plateStr)
63 plate = [plateList[0]]
64 b = [plateList[i][-1] for i in range(len(plateList))]
65 plate.extend(b[1:7])
66 return plateStr,plate
67
68 # 將生成的車牌圖片寫入文件夾，對應的label寫入label.txt
69 def genBatch(self, batchSize,pos,charRange, outputPath,size):
70 if (not os.path.exists(outputPath)):
71 os.mkdir(outputPath)
72 outfile = open('label.txt','w')
73 for i in xrange(batchSize):
74 plateStr,plate = G.genPlateString(-1,-1)
75 print plateStr,plate
76 img = G.generate(plateStr);
77 img = cv2.resize(img,size);
78 cv2.imwrite(outputPath + "/" + str(i).zfill(2) + ".jpg", img);
79 outfile.write(str(plate)+"\n")
80 G = GenPlate("./font/platech.ttf",'./font/platechar.ttf',"./NoPlates")
完整代碼：
View Code
運行時加生成數量和保存路徑即可，如：
1 python genPlate.py 100 ./plate_100
顯示結果：
上圖即為生成的車牌數據，有清晰的有模糊的，有比較方正的，也有一些比較傾斜，生成完大量的車牌樣張后就可以進行車牌識別了。下一小節將會講如何用端對端的CNN進行車牌識別，不需要通過傳統的ocr先對字符進行分割處理后再識別。
參考資料：
1.原來做的車牌識別項目：https://github.com/huxiaoman7/mxnet-cnn-plate-recognition
責任編輯：張燕妮
來源：
www.cnblogs.com
