OpenCV+TensorFlow图片手写数字识别(附源码)

100 阅读 0 评论 66 点赞

我是靠谱客的博主顺利路灯，这篇文章主要介绍OpenCV+TensorFlow图片手写数字识别(附源码)，现在分享给大家，希望可以做个参考。

初次接触TensorFlow，而手写数字训练识别是其最基本的入门教程，网上关于训练的教程很多，但是模型的测试大多都是官方提供的一些素材，能不能自己随便写一串数字让机器识别出来呢？纸上得来终觉浅，带着这个疑问昨晚研究了下，利用这篇文章来记录下自己的一些心得！

以下这个图片是我随机写的一串数字，我的目标是利用训练好的模型来识别出图片里面的手写数字，开始实战！

2层卷积神经网络的训练：

复制代码

from tensorflow.examples.tutorials.mnist import input_data# 保存模型需要的库from tensorflow.python.framework.graph_util import convert_variables_to_constantsfrom tensorflow.python.framework import graph_util# 导入其他库import tensorflow as tfimport cv2import numpy as np
# 获取MINIST数据mnist = input_data.read_data_sets("MNIST_data", one_hot=True)# 创建会话sess = tf.InteractiveSession()# 占位符x = tf.placeholder("float", shape=[None, 784], name="Mul")y_ = tf.placeholder("float", shape=[None, 10], name="y_")# 变量W = tf.Variable(tf.zeros([784, 10]), name='x')b = tf.Variable(tf.zeros([10]), 'y_')

# 权重def weight_variable(shape):    initial = tf.truncated_normal(shape, stddev=0.1)    return tf.Variable(initial)

# 偏差def bias_variable(shape):    initial = tf.constant(0.1, shape=shape)    return tf.Variable(initial)

# 卷积def conv2d(x, W):    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

# 最大池化def max_pool_2x2(x):    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],                          strides=[1, 2, 2, 1], padding='SAME')

# 相关变量的创建W_conv1 = weight_variable([5, 5, 1, 32])b_conv1 = bias_variable([32])x_image = tf.reshape(x, [-1, 28, 28, 1])h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)h_pool1 = max_pool_2x2(h_conv1)W_conv2 = weight_variable([5, 5, 32, 64])b_conv2 = bias_variable([64])# 激活函数h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)h_pool2 = max_pool_2x2(h_conv2)W_fc1 = weight_variable([7 * 7 * 64, 1024])b_fc1 = bias_variable([1024])W_fc2 = weight_variable([1024, 10])b_fc2 = bias_variable([10])h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)keep_prob = tf.placeholder("float", name='rob')h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)# 用于训练用的softmax函数y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2, name='res')# 用于训练作完后，作测试用的softmax函数y_conv2 = tf.nn.softmax(tf.matmul(h_fc1, W_fc2) + b_fc2, name="final_result")# 交叉熵的计算，返回包含了损失值的Tensor。cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))# 优化器，负责最小化交叉熵train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))# 计算准确率accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))# 初始化所以变量sess.run(tf.global_variables_initializer())# 保存输入输出，可以为之后用tf.add_to_collection('res', y_conv)tf.add_to_collection('output', y_conv2)tf.add_to_collection('x', x)# 训练开始for i in range(10000):    batch = mnist.train.next_batch(100)    if i % 100 == 0:        train_accuracy = accuracy.eval(feed_dict={            x: batch[0], y_: batch[1], keep_prob: 1.0})        print("step %d, training accuracy %g" % (i, train_accuracy))    # run()可以看做输入相关值给到函数中的占位符，然后计算的出结果，这里将batch[0]，给xbatch[1]给y_    train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})# 将当前图设置为默认图graph_def = tf.get_default_graph().as_graph_def()# 将上面的变量转化成常量，保存模型为pb模型时需要,注意这里的final_result和前面的y_con2是同名，只有这样才会保存它，否则会报错，# 如果需要保存其他tensor只需要让tensor的名字和这里保持一直即可output_graph_def = tf.graph_util.convert_variables_to_constants(sess, graph_def, ['final_result'])# 用saver 保存模型saver = tf.train.Saver()saver.save(sess, "model/model")

1
2
3
4
5
6
7
8
9
10
11
12
13
from tensorflow.examples.tutorials.mnist import input_data# 保存模型需要的库from tensorflow.python.framework.graph_util import convert_variables_to_constantsfrom tensorflow.python.framework import graph_util# 导入其他库import tensorflow as tfimport cv2import numpy as np
# 获取MINIST数据mnist = input_data.read_data_sets("MNIST_data", one_hot=True)# 创建会话sess = tf.InteractiveSession()# 占位符x = tf.placeholder("float", shape=[None, 784], name="Mul")y_ = tf.placeholder("float", shape=[None, 10], name="y_")# 变量W = tf.Variable(tf.zeros([784, 10]), name='x')b = tf.Variable(tf.zeros([10]), 'y_')

# 权重def weight_variable(shape):    initial = tf.truncated_normal(shape, stddev=0.1)    return tf.Variable(initial)

# 偏差def bias_variable(shape):    initial = tf.constant(0.1, shape=shape)    return tf.Variable(initial)

# 卷积def conv2d(x, W):    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

# 最大池化def max_pool_2x2(x):    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],                          strides=[1, 2, 2, 1], padding='SAME')

# 相关变量的创建W_conv1 = weight_variable([5, 5, 1, 32])b_conv1 = bias_variable([32])x_image = tf.reshape(x, [-1, 28, 28, 1])h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)h_pool1 = max_pool_2x2(h_conv1)W_conv2 = weight_variable([5, 5, 32, 64])b_conv2 = bias_variable([64])# 激活函数h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)h_pool2 = max_pool_2x2(h_conv2)W_fc1 = weight_variable([7 * 7 * 64, 1024])b_fc1 = bias_variable([1024])W_fc2 = weight_variable([1024, 10])b_fc2 = bias_variable([10])h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)keep_prob = tf.placeholder("float", name='rob')h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)# 用于训练用的softmax函数y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2, name='res')# 用于训练作完后，作测试用的softmax函数y_conv2 = tf.nn.softmax(tf.matmul(h_fc1, W_fc2) + b_fc2, name="final_result")# 交叉熵的计算，返回包含了损失值的Tensor。cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))# 优化器，负责最小化交叉熵train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))# 计算准确率accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))# 初始化所以变量sess.run(tf.global_variables_initializer())# 保存输入输出，可以为之后用tf.add_to_collection('res', y_conv)tf.add_to_collection('output', y_conv2)tf.add_to_collection('x', x)# 训练开始for i in range(10000):    batch = mnist.train.next_batch(100)    if i % 100 == 0:        train_accuracy = accuracy.eval(feed_dict={            x: batch[0], y_: batch[1], keep_prob: 1.0})        print("step %d, training accuracy %g" % (i, train_accuracy))    # run()可以看做输入相关值给到函数中的占位符，然后计算的出结果，这里将batch[0]，给xbatch[1]给y_    train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})# 将当前图设置为默认图graph_def = tf.get_default_graph().as_graph_def()# 将上面的变量转化成常量，保存模型为pb模型时需要,注意这里的final_result和前面的y_con2是同名，只有这样才会保存它，否则会报错，# 如果需要保存其他tensor只需要让tensor的名字和这里保持一直即可output_graph_def = tf.graph_util.convert_variables_to_constants(sess, graph_def, ['final_result'])# 用saver 保存模型saver = tf.train.Saver()saver.save(sess, "model/model")

网络训练成功后在model_data文件夹里有如下四个文件：

网络模型的验证可大致从以下三个部分来进行：

接下来就是要利用上面的图片来测试我们的模型。实际上图像的预处理部分很关键，也就是如何准确的提取出上面图像中的数字的区域，并且进行阈值分割，传统的单一阈值分割很难达到要求，因此本次分割采用基于改进的Niblack的分割方法，大家有兴趣可以查阅相关的资料。

分割完了之后要标记连通区域，去除那些小点区域。找到其外接矩形，可认为这个矩形区域就是我们感兴趣的区域。

降采样为28*28的大小来进行识别。

代码部分如下所示：

复制代码

import numpy as npimport cv2import matplotlib.pyplot as pltimport imutilsimport matplotlib.patches as mpatchesfrom skimage import data, segmentation, measure, morphology, colorimport tensorflow as tf

class Number_recognition():    """ 模型恢复初始化"""
    def __init__(self, img):        self.sess = tf.InteractiveSession()        saver = tf.train.import_meta_graph('model/model.meta')        saver.restore(self.sess, 'model/model')  # 模型恢复        # graph = tf.get_default_graph()        # 获取输入tensor,,获取输出tensor        self.input_x = self.sess.graph.get_tensor_by_name("Mul:0")        self.y_conv2 = self.sess.graph.get_tensor_by_name("final_result:0")        self.Preprocessing(img)  # 图像预处理
    def recognition(self, im):        im = cv2.resize(im, (28, 28), interpolation=cv2.INTER_CUBIC)        x_img = np.reshape(im, [-1, 784])        output = self.sess.run(self.y_conv2, feed_dict={self.input_x: x_img})        print('您输入的数字是 %d' % (np.argmax(output)))        return np.argmax(output)  # 返回识别的结果
    def Preprocessing(self, image):        if image.shape[0] > 1000:            image = imutils.resize(image, height=800)  # 如果图像太大局部阈值分割速度会稍慢些，因此图像太大时进行降采样
        img = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)  # convert to gray picture        m1, n1 = img.shape        k = int(m1 / 19) + 1        l = int(n1 / 19) + 1        #img = cv2.GaussianBlur(img, (3, 3), 0)  # 高斯滤波        imm = img.copy()        # 基于Niblack的局部阈值分割法，对于提取文本类图像分割效果比较好        for x in range(k):            for y in range(l):                s = imm[19 * x:19 * (x + 1), 19 * y:19 * (y + 1)]                me = s.mean()  # 均值                var = np.std(s)  # 方差                t = me * (1 - 0.2 * ((125 - var) / 125))                ret, imm[19 * x:19 * (x + 1), 19 * y:19 * (y + 1)] = cv2.threshold(                    imm[19 * x:19 * (x + 1), 19 * y:19 * (y + 1)], t, 255, cv2.THRESH_BINARY_INV)        label_image = measure.label(imm)  # 连通区域标记        for region in measure.regionprops(label_image):  # 循环得到每一个连通区域属性集            # 忽略小区域            if region.area < 100:                continue            minr, minc, maxr, maxc = region.bbox  # 得到外包矩形参数            cv2.rectangle(image, (minc, minr), (maxc, maxr), (0, 255, 0), 2)  # 绘制连通区域            im2 = imm[minr - 15:maxr + 15, minc - 15:maxc + 15]  # 获得感兴趣区域，也即每个数字的区
            number = self.recognition(im2)  # 进行识别            cv2.putText(image, str(number), (minc, minr - 10), 0, 2, (0, 0, 255), 2)  # 将识别结果写在原图上        cv2.imshow("threshold", imm)        cv2.imshow("recgonize result", image)        cv2.waitKey(0)

if __name__ == '__main__':    img = cv2.imread("./imgs/1.png")    x = Number_recognition(img)

1
2
3
4
5
6
7
8
9
10
import numpy as npimport cv2import matplotlib.pyplot as pltimport imutilsimport matplotlib.patches as mpatchesfrom skimage import data, segmentation, measure, morphology, colorimport tensorflow as tf

class Number_recognition():    """ 模型恢复初始化"""
    def __init__(self, img):        self.sess = tf.InteractiveSession()        saver = tf.train.import_meta_graph('model/model.meta')        saver.restore(self.sess, 'model/model')  # 模型恢复        # graph = tf.get_default_graph()        # 获取输入tensor,,获取输出tensor        self.input_x = self.sess.graph.get_tensor_by_name("Mul:0")        self.y_conv2 = self.sess.graph.get_tensor_by_name("final_result:0")        self.Preprocessing(img)  # 图像预处理
    def recognition(self, im):        im = cv2.resize(im, (28, 28), interpolation=cv2.INTER_CUBIC)        x_img = np.reshape(im, [-1, 784])        output = self.sess.run(self.y_conv2, feed_dict={self.input_x: x_img})        print('您输入的数字是 %d' % (np.argmax(output)))        return np.argmax(output)  # 返回识别的结果
    def Preprocessing(self, image):        if image.shape[0] > 1000:            image = imutils.resize(image, height=800)  # 如果图像太大局部阈值分割速度会稍慢些，因此图像太大时进行降采样
        img = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)  # convert to gray picture        m1, n1 = img.shape        k = int(m1 / 19) + 1        l = int(n1 / 19) + 1        #img = cv2.GaussianBlur(img, (3, 3), 0)  # 高斯滤波        imm = img.copy()        # 基于Niblack的局部阈值分割法，对于提取文本类图像分割效果比较好        for x in range(k):            for y in range(l):                s = imm[19 * x:19 * (x + 1), 19 * y:19 * (y + 1)]                me = s.mean()  # 均值                var = np.std(s)  # 方差                t = me * (1 - 0.2 * ((125 - var) / 125))                ret, imm[19 * x:19 * (x + 1), 19 * y:19 * (y + 1)] = cv2.threshold(                    imm[19 * x:19 * (x + 1), 19 * y:19 * (y + 1)], t, 255, cv2.THRESH_BINARY_INV)        label_image = measure.label(imm)  # 连通区域标记        for region in measure.regionprops(label_image):  # 循环得到每一个连通区域属性集            # 忽略小区域            if region.area < 100:                continue            minr, minc, maxr, maxc = region.bbox  # 得到外包矩形参数            cv2.rectangle(image, (minc, minr), (maxc, maxr), (0, 255, 0), 2)  # 绘制连通区域            im2 = imm[minr - 15:maxr + 15, minc - 15:maxc + 15]  # 获得感兴趣区域，也即每个数字的区
            number = self.recognition(im2)  # 进行识别            cv2.putText(image, str(number), (minc, minr - 10), 0, 2, (0, 0, 255), 2)  # 将识别结果写在原图上        cv2.imshow("threshold", imm)        cv2.imshow("recgonize result", image)        cv2.waitKey(0)

if __name__ == '__main__':    img = cv2.imread("./imgs/1.png")    x = Number_recognition(img)