[機器視覺]使用python自動識別驗證碼詳解

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

# 獲取驗證碼的網址
CAPT_URL = "http://xxxxxxxxxxxx.cn/servlet/ImageServlet"

# 驗證碼的保存路徑
CAPT_PATH = "capt/"
if not os.path.exists(CAPT_PATH):
  os.mkdir(CAPT_PATH)

# 將驗證碼轉為灰度圖時用到的"lookup table"
THRESHOLD = 165
LUT = [0]*THRESHOLD + [1]*(256 - THRESHOLD)

def capt_fetch():
  """
  從網站獲取驗證碼，將驗證碼轉為Image對象

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

  :param:
  :return capt: 一個Image對象
  """
  # 從網站獲取驗證碼
  capt_raw = requests.get(CAPT_URL)

  # 將二進制的驗證碼圖片寫入IO流
  f = BytesIO(capt_raw.content)

  # 將驗證碼轉換為Image對象
  capt = Image.open(f)

  return capt

def capt_download():
  """
  將Image類型的驗證碼對象保存到本地

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

  :require capt_fetch(): 從nbsc網站獲取驗證碼
  :require CAPT_PATH: 驗證碼保存路徑

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

  text = raw_input("請輸入驗證碼中的字符：")
  suffix = str(int(time.time() * 1e3))

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

def capt_process(capt):
  """
  圖像預處理：將驗證碼圖片轉為二值型圖片，按字符切割

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

  :param capt: 驗證碼Image對象
  :return capt_per_char_list: 一個數組包含四個元素，每個元素是一張包含單個字符的二值型圖片
  """
  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):
  """
  提取圖像特征

  :require numpy: import numpy as np

  :param capt_per_char: 由單個字符組成的二值型圖片
  :return char_features:一個數組，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():
  """
  將預分類的驗證碼圖片集轉化為字符特征訓練集

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

  :require capt_process(): 圖像預處理
  :require capt_inference(): 提取圖像特征

  :param:
  :return train_table: 驗證碼字符特征訓練集
  :return train_labels: 驗證碼字符預分類結果
  """
  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（近鄰分類法），
  根據已知特征矩陣的分類情況，預測未分類的特征向量所屬類別

  :require numpy: import numpy as np

  :param train_table: 預分類的特征矩陣
  :param test_vec: 特征向量， 長度必須與矩陣的列數相等
  :param labels: 特征矩陣的類別向量
  :return : 預測特征向量所屬的類別 
  """

  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():
  """
  測試模型分類效果

  :require Image: from PIL import Image

  :require capt_fetch(): 從nbsc網站獲取驗證碼
  :require capt_process(): 圖像預處理
  :require capt_inference(): 提取圖像特征
  :train_table, train_labels: train_table, train_labels = train()

  :param:
  :return capt: 驗證碼圖片
  :return test_labels: 驗證碼識別結果
  """

  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

# 下載120張圖片到本地
for i in range(120):
  capt_download()

# 模型的訓練與測試
train_table, train_labels = train()
test_capt, test_labels = test()