[视觉识别]OpenCV + CNN 大神符识别

数据集

Mnist数据集:http://yann.lecun.com/exdb/mnist/

训练

 import numpy as np
 from keras.datasets import mnist
 from keras.models import Sequential
 from keras.layers import Dense,Activation,Conv2D,MaxPooling2D,Flatten,Dropout,BatchNormalization,ZeroPadding2D
 from keras.optimizers import Adam
 from keras import backend as K
 K.backend()
 import tensorflow as tf
 import cv2 as cv
 import pandas as pd
 #加载数据
 datas=pd.read_csv("mnist_train.csv")
 images=datas.iloc[:,1:].values
 x_image=images.astype(np.float)
 x_image=np.multiply(x_image,1.0/255.0)
 labels=datas.iloc[:,0].values
 x_image[0]
 labels[0]
 #CNN模型加载
 def loadCNN():
     model = Sequential()
     model.add(Conv2D(32,(3,3),padding="valid",input_shape=(28,28,1)))
     convout1 = Activation("relu")
     model.add(convout1)
     model.add(BatchNormalization(epsilon=1e-6,axis=1))
     model.add(MaxPooling2D(pool_size=(2,2)))
     model.add(ZeroPadding2D((1,1)))
     model.add(Conv2D(48,(3,3)))
     convout2 = Activation("relu")
     model.add(convout2)
     model.add(BatchNormalization(epsilon=1e-6,axis=1))
     model.add(MaxPooling2D(pool_size=(2,2)))
     model.add(Conv2D(64,(2,2)))
     convout3 = Activation("relu")
     model.add(convout3)
     model.add(BatchNormalization(epsilon=1e-6,axis=1))
     model.add(MaxPooling2D(pool_size=(2,2)))
     model.add(Dropout(0.5))
     model.add(Flatten())
     model.add(Dense(3168))
     model.add(Activation("relu"))
     model.add(Dense(10))
     model.add(Activation("softmax"))
     model.compile(loss="categorical_crossentropy",optimizer="adam",metrics=['accuracy'])
     model.summary()
     model.get_config()
     return model
 #训练加存储
 from keras.utils import to_categorical
 x_input = x_image.reshape([-1,28,28,1])
 y_input = to_categorical(labels, num_classes=10)
 print(y_input.shape)
 model = loadCNN()
 hist = model.fit(x_input,y_input,batch_size = 32,epochs=15,verbose=1, validation_split=0.2)
 model.save_weights("model/mnist.hdf5",overwrite=True)

 Layer (type)                 Output Shape              Param #
 =================================================================
 conv2d_1 (Conv2D)            (None, 26, 26, 32)        320
 _________________________________________________________________
 activation_1 (Activation)    (None, 26, 26, 32)        0
 _________________________________________________________________
 batch_normalization_1 (Batch (None, 26, 26, 32)        104
 _________________________________________________________________
 max_pooling2d_1 (MaxPooling2 (None, 13, 13, 32)        0
 _________________________________________________________________
 zero_padding2d_1 (ZeroPaddin (None, 15, 15, 32)        0
 _________________________________________________________________
 conv2d_2 (Conv2D)            (None, 13, 13, 48)        13872
 _________________________________________________________________
 activation_2 (Activation)    (None, 13, 13, 48)        0
 _________________________________________________________________
 batch_normalization_2 (Batch (None, 13, 13, 48)        52
 _________________________________________________________________
 max_pooling2d_2 (MaxPooling2 (None, 6, 6, 48)          0
 _________________________________________________________________
 conv2d_3 (Conv2D)            (None, 5, 5, 64)          12352
 _________________________________________________________________
 activation_3 (Activation)    (None, 5, 5, 64)          0
 _________________________________________________________________
 batch_normalization_3 (Batch (None, 5, 5, 64)          20
 _________________________________________________________________
 max_pooling2d_3 (MaxPooling2 (None, 2, 2, 64)          0
 _________________________________________________________________
 dropout_1 (Dropout)          (None, 2, 2, 64)          0
 _________________________________________________________________
 flatten_1 (Flatten)          (None, 256)               0
 _________________________________________________________________
 dense_1 (Dense)              (None, 3168)              814176
 _________________________________________________________________
 activation_4 (Activation)    (None, 3168)              0
 _________________________________________________________________
 dense_2 (Dense)              (None, 10)                31690
 _________________________________________________________________
 activation_5 (Activation)    (None, 10)                0
 =================================================================
 Total params: 872,586
 Trainable params: 872,498
 Non-trainable params: 88
 _________________________________________________________________
 Train on 47999 samples, validate on 12000 samples
 Epoch 1/15
 47999/47999 [==============================] - 1641s 34ms/step - loss: 0.2381 - acc: 0.9250 - val_loss: 0.0724 - val_acc: 0.9762
 Epoch 2/15
 47999/47999 [==============================] - 1634s 34ms/step - loss: 0.1009 - acc: 0.9693 - val_loss: 0.0486 - val_acc: 0.9852
 Epoch 3/15
 47999/47999 [==============================] - 1626s 34ms/step - loss: 0.0787 - acc: 0.9768 - val_loss: 0.0472 - val_acc: 0.9863
 Epoch 4/15
 47999/47999 [==============================] - 1619s 34ms/step - loss: 0.0623 - acc: 0.9805 - val_loss: 0.0358 - val_acc: 0.9892
 Epoch 5/15
 47999/47999 [==============================] - 1627s 34ms/step - loss: 0.0568 - acc: 0.9829 - val_loss: 0.0427 - val_acc: 0.9878
 Epoch 6/15
 47999/47999 [==============================] - 1631s 34ms/step - loss: 0.0496 - acc: 0.9847 - val_loss: 0.0355 - val_acc: 0.9908
 Epoch 7/15
 47999/47999 [==============================] - 1626s 34ms/step - loss: 0.0430 - acc: 0.9871 - val_loss: 0.0284 - val_acc: 0.9922
 Epoch 8/15
 47999/47999 [==============================] - 1632s 34ms/step - loss: 0.0390 - acc: 0.9877 - val_loss: 0.0269 - val_acc: 0.9922
 Epoch 9/15
 47999/47999 [==============================] - 1632s 34ms/step - loss: 0.0363 - acc: 0.9886 - val_loss: 0.0341 - val_acc: 0.9904
 Epoch 10/15
 47999/47999 [==============================] - 1634s 34ms/step - loss: 0.0315 - acc: 0.9896 - val_loss: 0.0321 - val_acc: 0.9908
 Epoch 11/15
 47999/47999 [==============================] - 1626s 34ms/step - loss: 0.0301 - acc: 0.9907 - val_loss: 0.0325 - val_acc: 0.9912
 Epoch 12/15
 47999/47999 [==============================] - 1638s 34ms/step - loss: 0.0284 - acc: 0.9906 - val_loss: 0.0280 - val_acc: 0.9928
 Epoch 13/15
 47999/47999 [==============================] - 1635s 34ms/step - loss: 0.0261 - acc: 0.9920 - val_loss: 0.0313 - val_acc: 0.9919
 Epoch 14/15
 47999/47999 [==============================] - 1642s 34ms/step - loss: 0.0246 - acc: 0.9923 - val_loss: 0.0246 - val_acc: 0.9935
 Epoch 15/15
 47999/47999 [==============================] - 1639s 34ms/step - loss: 0.0228 - acc: 0.9926 - val_loss: 0.0288 - val_acc: 0.9922

测试

 def test():
     model = loadCNN()
     model.load_weights("model/mnist.hdf5")
     cap = cv.VideoCapture("test.wmv")
     while(cap.isOpened()):
         ret,frame = cap.read()
         img = draw(model,frame)
         cv.imshow(",img)
         if cv.waitKey(1) == 27:
             break
     cap.release()
     cv.destroyAllWindows()

 def draw(model,img):
     kernel = np.ones((3,3),np.uint8)
     gray = cv.cvtColor(img,cv.COLOR_BGR2GRAY)
     lower_hsv = np.array([0,0,223])
     upper_hsv = np.array([198,38,255])
     ROI = cv.cvtColor(img,cv.COLOR_BGR2HSV)
     frame = cv.inRange(ROI,lower_hsv,upper_hsv)
     image,contours, hierarchy= cv.findContours(frame,cv.RETR_LIST,cv.CHAIN_APPROX_SIMPLE)
     i=0
     for contour in contours:
         if len(contour) >3:
             x,y,w,h = cv.boundingRect(contour)
             if w/h>=1.35 and w>15 and h >20 and w <=45:
                 #print(x,y,w,h)
                 p_img = gray[y:y+h, x+w//2-h//2:x+w//2+h//2]
                 p_img = cv.erode(p_img,kernel)
                 p_img = cv.resize(p_img,(28,28),interpolation= cv.INTER_LINEAR)
                 ph,pw = p_img.shape
                 for hx in range(ph):
                     for wy in range(pw):
                         p_img[hx][wy] = 255- p_img[hx][wy]
                 p_img = np.array(p_img,'f')
                 p_img = p_img/ 255.0
                 p_img = p_img.reshape([-1,28,28,1])
                 pdt = np.argmax(model.predict(p_img))
                 cv.putText(img,str(pdt),(x,y),cv.FONT_HERSHEY_COMPLEX,0.8,(0,0,255),1)
                 ####这里用来判断
                 #print(x,y,w,h)
                 #cv.rectangle(img, (x+ w//2-h//2,y), (x+w//2+h//2, y + h), (0, 0, 225), 3)
     #cv.drawContours(img, contours, -1, (0, 0, 255), 3)
     #print(i)
     return img
 test()