在mnist_predict目录下新建文件，命名为predict_pic.py，识别图像。
import os
import cv2
import numpy as np
import tensorflow as tf
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

将输入的彩色图像转换为二值化图

def color_input(endimg):
img_gray = cv2.cvtColor(endimg, cv2.COLOR_BGR2GRAY) # 灰度化
ret, img_threshold = cv2.threshold(img_gray, 127, 255, cv2.THRESH_BINARY_INV)
return img_threshold

读取图像并显示

def read_pic(path):
img = cv2.imread(path, cv2.IMREAD_COLOR)
cv2.imshow('img', img)
cv2.waitKey(0)
img_threshold = color_input(img)
cv2.imshow('img_threshold', img_threshold)
cv2.waitKey(0)
return img_threshold
if name == 'main':
with tf.Session() as sess:
saver = tf.train.import_meta_graph('model_data/model.meta')

        # 模型恢复
        saver.restore(sess, 'model_data/model')
        graph = tf.get_default_graph()
        # 获取变量
        input_x = sess.graph.get_tensor_by_name("Mul:0")
        y_conv2 = sess.graph.get_tensor_by_name("final_result:0")
        # 读取图像
        img_threshold = read_pic("nine.png")
        # 将图像进行缩放
        im = cv2.resize(img_threshold, (28, 28), interpolation=cv2.INTER_CUBIC)
        x_img = np.reshape(im, [-1, 784])
        # 识别
        output = sess.run(y_conv2, feed_dict={input_x: x_img})
        result = np.argmax(output)
        print("识别结果为：{}".format(result))