OpenCV + CNN 大神符識別
阿新 • • 發佈：2018-10-07
rpo cat test oca pool output ros label entropy
數據集
Mnist數據集:http://yann.lecun.com/exdb/mnist/
訓練
1 import numpy as np 2 from keras.datasets import mnist 3 from keras.models import Sequential 4 from keras.layers import Dense,Activation,Conv2D,MaxPooling2D,Flatten,Dropout,BatchNormalization,ZeroPadding2D 5 from keras.optimizers import Adam 6 fromkeras import backend as K 7 K.backend() 8 import tensorflow as tf 9 import cv2 as cv 10 import pandas as pd 11 #加載數據 12 datas=pd.read_csv("mnist_train.csv") 13 images=datas.iloc[:,1:].values 14 x_image=images.astype(np.float) 15 x_image=np.multiply(x_image,1.0/255.0) 16 labels=datas.iloc[:,0].values17 x_image[0] 18 labels[0] 19 #CNN模型加載 20 def loadCNN(): 21 model = Sequential() 22 model.add(Conv2D(32,(3,3),padding="valid",input_shape=(28,28,1))) 23 convout1 = Activation("relu") 24 model.add(convout1) 25 model.add(BatchNormalization(epsilon=1e-6,axis=1)) 26 model.add(MaxPooling2D(pool_size=(2,2)))27 model.add(ZeroPadding2D((1,1))) 28 model.add(Conv2D(48,(3,3))) 29 convout2 = Activation("relu") 30 model.add(convout2) 31 model.add(BatchNormalization(epsilon=1e-6,axis=1)) 32 model.add(MaxPooling2D(pool_size=(2,2))) 33 model.add(Conv2D(64,(2,2))) 34 convout3 = Activation("relu") 35 model.add(convout3) 36 model.add(BatchNormalization(epsilon=1e-6,axis=1)) 37 model.add(MaxPooling2D(pool_size=(2,2))) 38 model.add(Dropout(0.5)) 39 model.add(Flatten()) 40 model.add(Dense(3168)) 41 model.add(Activation("relu")) 42 model.add(Dense(10)) 43 model.add(Activation("softmax")) 44 model.compile(loss="categorical_crossentropy",optimizer="adam",metrics=[‘accuracy‘]) 45 model.summary() 46 model.get_config() 47 return model 48 #訓練加存儲 49 from keras.utils import to_categorical 50 x_input = x_image.reshape([-1,28,28,1]) 51 y_input = to_categorical(labels, num_classes=10) 52 print(y_input.shape) 53 model = loadCNN() 54 hist = model.fit(x_input,y_input,batch_size = 32,epochs=15,verbose=1, validation_split=0.2) 55 model.save_weights("model/mnist.hdf5",overwrite=True)
1 Layer (type) Output Shape Param # 2 ================================================================= 3 conv2d_1 (Conv2D) (None, 26, 26, 32) 320 4 _________________________________________________________________ 5 activation_1 (Activation) (None, 26, 26, 32) 0 6 _________________________________________________________________ 7 batch_normalization_1 (Batch (None, 26, 26, 32) 104 8 _________________________________________________________________ 9 max_pooling2d_1 (MaxPooling2 (None, 13, 13, 32) 0 10 _________________________________________________________________ 11 zero_padding2d_1 (ZeroPaddin (None, 15, 15, 32) 0 12 _________________________________________________________________ 13 conv2d_2 (Conv2D) (None, 13, 13, 48) 13872 14 _________________________________________________________________ 15 activation_2 (Activation) (None, 13, 13, 48) 0 16 _________________________________________________________________ 17 batch_normalization_2 (Batch (None, 13, 13, 48) 52 18 _________________________________________________________________ 19 max_pooling2d_2 (MaxPooling2 (None, 6, 6, 48) 0 20 _________________________________________________________________ 21 conv2d_3 (Conv2D) (None, 5, 5, 64) 12352 22 _________________________________________________________________ 23 activation_3 (Activation) (None, 5, 5, 64) 0 24 _________________________________________________________________ 25 batch_normalization_3 (Batch (None, 5, 5, 64) 20 26 _________________________________________________________________ 27 max_pooling2d_3 (MaxPooling2 (None, 2, 2, 64) 0 28 _________________________________________________________________ 29 dropout_1 (Dropout) (None, 2, 2, 64) 0 30 _________________________________________________________________ 31 flatten_1 (Flatten) (None, 256) 0 32 _________________________________________________________________ 33 dense_1 (Dense) (None, 3168) 814176 34 _________________________________________________________________ 35 activation_4 (Activation) (None, 3168) 0 36 _________________________________________________________________ 37 dense_2 (Dense) (None, 10) 31690 38 _________________________________________________________________ 39 activation_5 (Activation) (None, 10) 0 40 ================================================================= 41 Total params: 872,586 42 Trainable params: 872,498 43 Non-trainable params: 88 44 _________________________________________________________________ 45 Train on 47999 samples, validate on 12000 samples 46 Epoch 1/15 47 47999/47999 [==============================] - 1641s 34ms/step - loss: 0.2381 - acc: 0.9250 - val_loss: 0.0724 - val_acc: 0.9762 48 Epoch 2/15 49 47999/47999 [==============================] - 1634s 34ms/step - loss: 0.1009 - acc: 0.9693 - val_loss: 0.0486 - val_acc: 0.9852 50 Epoch 3/15 51 47999/47999 [==============================] - 1626s 34ms/step - loss: 0.0787 - acc: 0.9768 - val_loss: 0.0472 - val_acc: 0.9863 52 Epoch 4/15 53 47999/47999 [==============================] - 1619s 34ms/step - loss: 0.0623 - acc: 0.9805 - val_loss: 0.0358 - val_acc: 0.9892 54 Epoch 5/15 55 47999/47999 [==============================] - 1627s 34ms/step - loss: 0.0568 - acc: 0.9829 - val_loss: 0.0427 - val_acc: 0.9878 56 Epoch 6/15 57 47999/47999 [==============================] - 1631s 34ms/step - loss: 0.0496 - acc: 0.9847 - val_loss: 0.0355 - val_acc: 0.9908 58 Epoch 7/15 59 47999/47999 [==============================] - 1626s 34ms/step - loss: 0.0430 - acc: 0.9871 - val_loss: 0.0284 - val_acc: 0.9922 60 Epoch 8/15 61 47999/47999 [==============================] - 1632s 34ms/step - loss: 0.0390 - acc: 0.9877 - val_loss: 0.0269 - val_acc: 0.9922 62 Epoch 9/15 63 47999/47999 [==============================] - 1632s 34ms/step - loss: 0.0363 - acc: 0.9886 - val_loss: 0.0341 - val_acc: 0.9904 64 Epoch 10/15 65 47999/47999 [==============================] - 1634s 34ms/step - loss: 0.0315 - acc: 0.9896 - val_loss: 0.0321 - val_acc: 0.9908 66 Epoch 11/15 67 47999/47999 [==============================] - 1626s 34ms/step - loss: 0.0301 - acc: 0.9907 - val_loss: 0.0325 - val_acc: 0.9912 68 Epoch 12/15 69 47999/47999 [==============================] - 1638s 34ms/step - loss: 0.0284 - acc: 0.9906 - val_loss: 0.0280 - val_acc: 0.9928 70 Epoch 13/15 71 47999/47999 [==============================] - 1635s 34ms/step - loss: 0.0261 - acc: 0.9920 - val_loss: 0.0313 - val_acc: 0.9919 72 Epoch 14/15 73 47999/47999 [==============================] - 1642s 34ms/step - loss: 0.0246 - acc: 0.9923 - val_loss: 0.0246 - val_acc: 0.9935 74 Epoch 15/15 75 47999/47999 [==============================] - 1639s 34ms/step - loss: 0.0228 - acc: 0.9926 - val_loss: 0.0288 - val_acc: 0.9922
測試
1 def test(): 2 model = loadCNN() 3 model.load_weights("model/mnist.hdf5") 4 cap = cv.VideoCapture("test.wmv") 5 while(cap.isOpened()): 6 ret,frame = cap.read() 7 img = draw(model,frame) 8 cv.imshow("2",img) 9 if cv.waitKey(1) == 27: 10 break 11 cap.release() 12 cv.destroyAllWindows() 13 14 15 def draw(model,img): 16 kernel = np.ones((3,3),np.uint8) 17 gray = cv.cvtColor(img,cv.COLOR_BGR2GRAY) 18 lower_hsv = np.array([0,0,223]) 19 upper_hsv = np.array([198,38,255]) 20 ROI = cv.cvtColor(img,cv.COLOR_BGR2HSV) 21 frame = cv.inRange(ROI,lower_hsv,upper_hsv) 22 image,contours, hierarchy= cv.findContours(frame,cv.RETR_LIST,cv.CHAIN_APPROX_SIMPLE) 23 i=0 24 for contour in contours: 25 if len(contour) >3: 26 x,y,w,h = cv.boundingRect(contour) 27 if w/h>=1.35 and w>15 and h >20 and w <=45: 28 #print(x,y,w,h) 29 p_img = gray[y:y+h, x+w//2-h//2:x+w//2+h//2] 30 p_img = cv.erode(p_img,kernel) 31 p_img = cv.resize(p_img,(28,28),interpolation= cv.INTER_LINEAR) 32 ph,pw = p_img.shape 33 for hx in range(ph): 34 for wy in range(pw): 35 p_img[hx][wy] = 255- p_img[hx][wy] 36 p_img = np.array(p_img,‘f‘) 37 p_img = p_img/ 255.0 38 p_img = p_img.reshape([-1,28,28,1]) 39 pdt = np.argmax(model.predict(p_img)) 40 cv.putText(img,str(pdt),(x,y),cv.FONT_HERSHEY_COMPLEX,0.8,(0,0,255),1) 41 ####這裏用來判斷 42 #print(x,y,w,h) 43 #cv.rectangle(img, (x+ w//2-h//2,y), (x+w//2+h//2, y + h), (0, 0, 225), 3) 44 #cv.drawContours(img, contours, -1, (0, 0, 255), 3) 45 #print(i) 46 return img 47 test()
OpenCV + CNN 大神符識別