Python机器视觉自动识别验证码实战：从预处理到近邻分类

import requests
import time
from io import BytesIO
from PIL import Image
import os
import numpy as np

# 获取验证码的网址
CAPT_URL = "http://xxxxxxxxxxxx.cn/servlet/ImageServlet"

# 验证码的保存路径
CAPT_PATH = "capt/"
if not os.path.exists(CAPT_PATH):
  os.mkdir(CAPT_PATH)

# 将验证码转为灰度图时用到的"lookup table"
THRESHOLD = 165
LUT = [0]*THRESHOLD + [1]*(256 - THRESHOLD)

def capt_fetch():
  """
  从网站获取验证码，将验证码转为Image对象

  :require requests: import requests
  :require time: import time
  :require BytesIO: from io import BytesIO
  :require Image: from PIL import Image

  :param:
  :return capt: 一个Image对象
  """
  # 从网站获取验证码
  capt_raw = requests.get(CAPT_URL)

  # 将二进制的验证码图片写入IO流
  f = BytesIO(capt_raw.content)

  # 将验证码转换为Image对象
  capt = Image.open(f)

  return capt

def capt_download():
  """
  将Image类型的验证码对象保存到本地

  :require Image: from PIL import Image
  :require os: import os

  :require capt_fetch(): 从nbsc网站获取验证码
  :require CAPT_PATH: 验证码保存路径

  :param:
  :return: 
  """
  capt = capt_fetch()
  capt.show()

  text = raw_input("请输入验证码中的字符：")
  suffix = str(int(time.time() * 1e3))

  capt.save(CAPT_PATH + text + "_" + suffix + ".jpg")

def capt_process(capt):
  """
  图像预处理：将验证码图片转为二值型图片，按字符切割

  :require Image: from PIL import Image
  :require LUT: A lookup table, 包含256个值

  :param capt: 验证码Image对象
  :return capt_per_char_list: 一个数组包含四个元素，每个元素是一张包含单个字符的二值型图片
  """
  capt_gray = capt.convert("L")
  capt_bw = capt_gray.point(LUT, "1")

  capt_per_char_list = []
  for i in range(4):
    x = 5 + i * 15
    y = 2
    capt_per_char = capt_bw.crop((x, y, x + 15, y + 18))
    capt_per_char_list.append(capt_per_char)

  return capt_per_char_list

def capt_inference(capt_per_char):
  """
  提取图像特征

  :require numpy: import numpy as np

  :param capt_per_char: 由单个字符组成的二值型图片
  :return char_features:一个数组，包含 capt_per_char中字符的特征
  """
  char_array = np.array(capt_per_char)

  total_pixels = np.sum(char_array)
  cols_pixels = np.sum(char_array, 0)
  rows_pixels = np.sum(char_array, 1)

  char_features = np.append(cols_pixels, rows_pixels)
  char_features = np.append(total_pixels, char_features)

  return char_features.tolist()

def train():
  """
  将预分类的验证码图片集转化为字符特征训练集

  :require Image: from PIL import Image
  :require os: import os

  :require capt_process(): 图像预处理
  :require capt_inference(): 提取图像特征

  :param:
  :return train_table: 验证码字符特征训练集
  :return train_labels: 验证码字符预分类结果
  """
  files = os.listdir(CAPT_PATH)

  train_table = []
  train_labels = []

  for f in files:
    train_labels += list(f.split("_")[0])

    capt = Image.open(CAPT_PATH + f)
    capt_per_char_list = capt_process(capt)
    for capt_per_char in capt_per_char_list:
      char_features = capt_inference(capt_per_char)
      train_table.append(char_features)

  return train_table, train_labels

def nnc(train_table, test_vec, train_labels):
  """
  Nearest Neighbour Classification（近邻分类法），
  根据已知特征矩阵的分类情况，预测未分类的特征向量所属类别

  :require numpy: import numpy as np

  :param train_table: 预分类的特征矩阵
  :param test_vec: 特征向量， 长度必须与矩阵的列数相等
  :param labels: 特征矩阵的类别向量
  :return : 预测特征向量所属的类别 
  """

  dist_mat = np.square(np.subtract(train_table, test_vec))
  dist_vec = np.sum(dist_mat, axis = 1)
  pos = np.argmin(dist_vec)

  return train_labels[pos]

def test():
  """
  测试模型分类效果

  :require Image: from PIL import Image

  :require capt_fetch(): 从nbsc网站获取验证码
  :require capt_process(): 图像预处理
  :require capt_inference(): 提取图像特征
  :train_table, train_labels: train_table, train_labels = train()

  :param:
  :return capt: 验证码图片
  :return test_labels: 验证码识别结果
  """

  test_labels = []

  capt = capt_fetch()
  capt_per_char_list = capt_process(capt)
  for capt_per_char in capt_per_char_list:
    char_features = capt_inference(capt_per_char)
    label = nnc(train_table, char_features, train_labels)
    test_labels.append(label)

  test_labels = "".join(test_labels)

  return capt, test_labels

# 下载120张图片到本地
for i in range(120):
  capt_download()

# 模型的训练与测试
train_table, train_labels = train()
test_capt, test_labels = test()