数字图像处理（名字的分割与识别）

Posted 2021-07-15 一只夫夫

内容

利用所学图像分割技术，将图片中包含的名字进行分割与识别

操作步骤

1. 利用基于边界的分割技术，提取图像边缘信息，然后根据灰度分布，提供数个可能存在名字的选区框
2. 训练一个能识别出图像中是否有名字的卷积神经网络
3. 去重后，若CNN识别到选区中有名字，则显示该选区框

1. 提取边缘，产生可能存在名字的选区框

import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
import PIL.Image as im
from function import*

dim = 256
img = plt.imread('2.jpg')
img = img/255.0
img = tf.image.resize(img, (dim,dim))
plt.title("original picture")
plt.imshow(img)

<matplotlib.image.AxesImage at 0x1968646a148>

# 使用log模板，对原图进行卷积，提取边缘
img_in = np.zeros((1, dim, dim, 1)) #首位1表示1张图片，末位1表示单通道
img_in[0,:,:,0] = img[:,:,0] #生成灰度图像
fil = np.zeros((5, 5, 1, 1)) #滤波器
fil[:,:,0,0]=[[0,  0, -1,  0, 0], 
   [0, -1, -2, -1, 0],
  [-1, -2, 16, -2, -1],
  [0, -1, -2, -1, 0],
  [0, 0, -1, 0, 0 ]]
img_out = tf.nn.conv2d(input = img_in, filters = fil, strides = 1, padding = 'SAME')
img_out = np.clip(img_out, 0, 1)
ccc = np.zeros((dim,dim,3))
ccc[:,:,0] = img_out[0,:,:,0]
ccc[:,:,1] = img_out[0,:,:,0]
ccc[:,:,2] = img_out[0,:,:,0]
plt.imshow(ccc)

# 产生可能存在名字的选区，并将这些“提议”选区显示出来
co = []
mat = img_out[0,:,:,0]
gray_mean = np.mean(mat)
gray_std = np.var(mat)
for x in range(50, 230, 30):
    for y in range(50, 230, 30):
        mean = np.mean(mat[x-50:x+50, y-50:y+50])
        if(mean - gray_mean > 0):
            co.append((x,y))
RGBmat = creat_RGBmat(ccc,co)
plt.figure(1)
plt.title("proposal blocks")
plt.imshow(RGBmat)

plt.figure(2)
picture_test = creat_X_test(img, co, 100)
print("已分割出选区数：",len(co))
print("block shape:",picture_test.shape)
for i in range(picture_test.shape[0]):
    plt.subplot(2,4,i+1)
    plt.imshow(picture_test[i])
    plt.axis('off')

2. 训练一个卷积神经网络

xx,yy = load_picture()
xx_train, yy_train, xx_test, yy_test = pre_process(xx, yy, 6)
xx_train = tf.image.resize(xx_train, (100,100))
xx_test = tf.image.resize(xx_test, (100,100))
print("xx_trian.shape:",xx_train.shape)
print("xx_test.shape:",xx_test.shape)
print("\\t\\t显示6张用于训练的图片")
for i in range(6):
    plt.subplot(2,3,i+1)
    plt.title(yy_train[i])
    plt.imshow(xx_train[i])
    plt.axis('off')

# 定义一个模型
def model_init_fun(input_shape, num_classes):
    initializer = tf.initializers.VarianceScaling(scale=2.0)
    layers = ([
        tf.keras.layers.Conv2D(input_shape = input_shape, filters = 30, padding = 'same', 
                               kernel_size = (10,10), activation = 'relu',kernel_initializer=initializer, name = 'conv1'),

        tf.keras.layers.Conv2D(filters = 60, padding = 'same', kernel_size = (5,5), 
                               activation = 'relu', kernel_initializer=initializer, name = 'conv2'),
        tf.keras.layers.Conv2D(filters = 120, padding = 'same', kernel_size = (3,3), 
                       activation = 'relu', kernel_initializer=initializer, name = 'conv3'),
        tf.keras.layers.MaxPool2D(pool_size = (4,4)),
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(num_classes, kernel_initializer=initializer, name = 'fc'),
        tf.keras.layers.Softmax()
    ])
    model = tf.keras.Sequential(layers)
    return model

def optimizer_init_fn(learning_rate):
    return tf.keras.optimizers.Adam(learning_rate=learning_rate) 

cnn = model_init_fun((100, 100, 3), 2)
opt = optimizer_init_fn(1e-4)

# 训练模型并显示部分验证结果
cnn.compile(optimizer = opt, loss = 'sparse_categorical_crossentropy',
                    metrics=[tf.keras.metrics.sparse_categorical_accuracy])
cnn.fit(x = xx_train, y = yy_train, epochs=20, verbose=0)
cnn.evaluate(x = xx_test, y = yy_test,verbose=2)

y_pre = cnn.predict_classes(xx_test)
print("\\n\\t\\t\\t验证结果")
for i in range(xx_test.shape[0]):
    plt.subplot(2, xx_test.shape[0]/2, i+1)
    if y_pre[i] == 1:
        title = "detected"
    else:
            title = "no"
    plt.title(title)
    plt.imshow(xx_test[i])
    plt.axis('off')

6/6 - 0s - loss: 0.0607 - sparse_categorical_accuracy: 1.0000

			验证结果

3. 对选区进行检测，显示检测到名字的选区框

# 逐个检测选区
y_pre = cnn.predict_classes(picture_test)
print("\\n\\t\\t\\t检测结果")
for i in range(y_pre.shape[0]):
    plt.subplot(3, 4, i+1)
    if y_pre[i] == 1:
        title = "detected"
    else:
            title = "no"
    plt.title(title)
    plt.imshow(picture_test[i])
    plt.axis('off')

			检测结果

#对选区去重并显示
co_final = []
index = np.where(y_pre == 1)
index = index[0]
for i in range(len(index)):
        x,y = co[index[i]]
        if(i==len(index)-1):
            px,py=co[i-1]
            dist = (x-px)**2 + (y-py)**2
            if(dist > 50**2):
                co_final.append((x,y))
        else:
            px,py=co[i+1]
            co_final.append((x,y))
            dist = (x-px)**2 + (y-py)**2
            if(dist < 50**2):
                i = i + 1
                
final_picture = creat_RGBmat(np.array(img), co_final)
plt.title("picture detected")
plt.imshow(final_picture)                

<matplotlib.image.AxesImage at 0x22244d0edc8>