KNN算法实战:验证码的识别

识别验证码的方式很多,如tesseract、SVM等。前面的几篇文章介绍了KNN算法,今天主要学习的是如何使用KNN进行验证码的识别。

数据准备

本次实验采用的是CSDN的验证码做演练,相关的接口:https://download.csdn.net/index.php/rest/tools/validcode/source_ip_validate/10.5711163911089325

目前接口返回的验证码共2种:

  • 《KNN算法实战:验证码的识别》  纯数字、干扰小的验证码,简单进行图片去除背景、二值化和阈值处理后,使用kNN算法即可识别。
  • 《KNN算法实战:验证码的识别》 字母加数字、背景有干扰、图形字符位置有轻微变形,进行图片去除背景、二值化和阈值处理后,使用kNN算法识别

这里选择第二种进行破解。由于两种验证码的图片大小不一样,所以可以使用图片大小来判断哪个是第一种验证码,哪个是第二种验证码。

下载验证码

import requests
import uuid
from PIL import Image
import os
url = "http://download.csdn.net/index.php/rest/tools/validcode/source_ip_validate/10.5711163911089325"
for i in range(1000):
    resp = requests.get(url)
    filename = "./captchas/" + str(uuid.uuid4()) + ".png"
    with open(filename, 'wb') as f:
        for chunk in resp.iter_content(chunk_size=1024):
            if chunk:  # filter out keep-alive new chunks
                f.write(chunk)
                f.flush()
        f.close()
    im = Image.open(filename)
    if im.size != (70, 25):
        im.close()
        os.remove(filename)
    else:
        print(filename)

分割字符

下载过后,就需要对字母进行分割。分割字符还是一件比较麻烦的工作。

灰度化

将彩色的图片转化为灰度图片,便于后面的二值化处理,示例代码:

from PIL import Image
 
file = ".\\captchas\\0a4a22cd-f16b-4ae4-bc52-cdf4c081301d.png"
im = Image.open(file)
im_gray = im.convert('L')
im_gray.show()

处理前:《KNN算法实战:验证码的识别》

处理后:《KNN算法实战:验证码的识别》

二值化

灰度化以后,有颜色的像素点为0-255之间的值。二值化就是将大于某个值的像素点都修改为255,小于该值的修改为0,示例代码:

from PIL import Image
import numpy as np
file = ".\\captchas\\0a4a22cd-f16b-4ae4-bc52-cdf4c081301d.png"
im = Image.open(file)
im_gray = im.convert('L')
# im_gray.show()
 
pix = np.array(im_gray)
print(pix.shape)
print(pix)
 
threshold = 100 #阈值
 
pix = (pix > threshold) * 255
print(pix)
 
out = Image.fromarray(pix)
out.show()

二值化输出的结果:《KNN算法实战:验证码的识别》

去除边框

从二值化输出的结果可以看到除了字符,还存在边框,在切割字符前还需要先将边框去除。

border_width = 1
new_pix = pix[border_width:-border_width,border_width:-border_width]

字符切割

由于字符与字符间没有存在连接,可以使用比较简单的“投影法”进行字符的切割。原理就是将二值化后的图片先在垂直方向进行投影,根据投影后的极值来判断分割边界。分割后的小图片再在水平方向进行投影。

《KNN算法实战:验证码的识别》

代码实现:

def vertical_image(image):
    height, width = image.shape
    h = [0] * width
    for x in range(width):
        for y in range(height):
            s = image[y, x]
            if s == 255:
                h[x] += 1
    new_image = np.zeros(image.shape, np.uint8)
    for x in range(width):
        cv2.line(new_image, (x, 0), (x, h[x]), 255, 1)
    cv2.imshow('vert_image', new_image)
    cv2.waitKey()
cv2.destroyAllWindows()

整体代码

from PIL import Image
import cv2
import numpy as np
import os
import uuid
 
 
def clean_bg(filename):
    im = Image.open(filename)
    im_gray = im.convert('L')
    image = np.array(im_gray)
    threshold = 100  # 阈值
    pix = (image > threshold) * 255
    border_width = 1
    new_image = pix[border_width:-border_width, border_width:-border_width]
    return new_image
 
 
def get_col_rect(image):
    height, width = image.shape
    h = [0] * width
    for x in range(width):
        for y in range(height):
            s = image[y, x]
            if s == 0:
                h[x] += 1
    col_rect = []
    in_line = False
    start_line = 0
    blank_distance = 1
    for i in range(len(h)):
        if not in_line and h[i] >= blank_distance:
            in_line = True
            start_line = i
        elif in_line and h[i] < blank_distance:
            rect = (start_line, i)
            col_rect.append(rect)
            in_line = False
            start_line = 0
    return col_rect
 
 
def get_row_rect(image):
    height, width = image.shape
    h = [0] * height
    for y in range(height):
        for x in range(width):
            s = image[y, x]
            if s == 0:
                h[y] += 1
    in_line = False
    start_line = 0
    blank_distance = 1
    row_rect = (0, 0)
    for i in range(len(h)):
        if not in_line and h[i] >= blank_distance:
            in_line = True
            start_line = i
        elif in_line and i == len(h)-1:
            row_rect = (start_line, i)
        elif in_line and h[i] < blank_distance:
            row_rect = (start_line, i)
            break
    return row_rect
 
 
def get_block_image(image, col_rect):
    col_image = image[0:image.shape[0], col_rect[0]:col_rect[1]]
    row_rect = get_row_rect(col_image)
    if row_rect[1] != 0:
        block_image = image[row_rect[0]:row_rect[1], col_rect[0]:col_rect[1]]
    else:
        block_image = None
    return block_image
 
 
def clean_bg(filename):
    im = Image.open(filename)
    im_gray = im.convert('L')
    image = np.array(im_gray)
    threshold = 100  # 阈值
    pix = (image > threshold) * 255
    border_width = 2
    new_image = pix[border_width:-border_width, border_width:-border_width]
    return new_image
 
def split(filename):
    image = clean_bg(filename)
    col_rect = get_col_rect(image)
    for cols in col_rect:
        block_image = get_block_image(image, cols)
        if block_image is not None:
            new_image_filename = 'letters/' + str(uuid.uuid4()) + '.png'
            cv2.imwrite(new_image_filename, block_image)
 
 
if __name__ == '__main__':
    for filename in os.listdir('captchas'):
        current_file = 'captchas/' + filename
        split(current_file)
        print('split file:%s' % current_file)

数据集准备

在完成图像切割后,需要做将切分的字母建立由标签的样本。即将切分后的字符梳理到正确的分类中。比较常见的方式是人工梳理。

由于图像比较多,这里使用使用Tesseract-OCR进行识别。

官方项目地址:github.com/tesseract-o…

Windows安装包地址:github.com/UB-Mannheim…

Tesseract-OCR的安装

下载完安装包后,直接运行安装即可,比较重要的是环境变量的设置。

  • 将安装目录(D:\Program Files (x86)\Tesseract-OCR)添加进PATH
  • 新建TESSDATA_PREFIX系统变量,值为tessdata 文件夹的路径(D:\Program Files (x86)\Tesseract-OCR\tessdata)
  • 安装Python包pytesseract(pip install pytesseract)

Tesseract-OCR的使用

使用起来非常的简单,代码如下:

from PIL import Image
import pytesseract
import os
 
 
def copy_to_dir(filename):
    image = Image.open(filename)
    code = pytesseract.image_to_string(image, config="-c tessedit"
                                                     "_char_whitelist=ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789"
                                                     " --psm 10"
                                                     " -l osd"
                                                     " ")
    if not os.path.exists("dataset/" + code):
        os.mkdir("dataset/" + code)
    image.save("dataset/" + code + filename.replace("letters", ""))
    image.close()
 
 
if __name__ == "__main__":
    for filename in os.listdir('letters'):
        current_file = 'letters/' + filename
        copy_to_dir(current_file)
        print(current_file)

由于Tesseract-OCR识别的准确率非常的低,完全不能使用,放弃~,还是需要手工整理。

图片尺寸统一

在完成人工处理后,发现切割后的图片大小不一。在字符识别前需要对图片进行的尺寸进行统一。

具体实现方法:

import cv2
 
def image_resize(filename):
    img = cv2.imread(filename, cv2.IMREAD_GRAYSCALE) #读取图片时采用单通道
    print(img)
    if img.shape[0] != 10 or img.shape[1] != 6:
        img = cv2.resize(img, (6, 10), interpolation=cv2.INTER_CUBIC)
        print(img)
        cv2.imwrite(filename, img)

使用cv2.resize时,参数输入是 宽×高×通道,这里使用的时单通道的,interpolation的选项有:

  • INTER_NEAREST 最近邻插值
  • INTER_LINEAR 双线性插值(默认设置)
  • INTER_AREA 使用像素区域关系进行重采样。 它可能是图像抽取的首选方法,因为它会产生无云纹理的结果。 但是当图像缩放时,它类似于INTER_NEAREST方法。
  • INTER_CUBIC 4×4像素邻域的双三次插值
  • INTER_LANCZOS4 8×8像素邻域的Lanczos插值

另外为了让数据更加便于利用,可以将图片再进行二值化的归一。具体代码如下:

import cv2
import numpy as np
 
def image_normalize(filename):
    img = cv2.imread(filename, cv2.IMREAD_GRAYSCALE) #读取图片时采用单通道
    if img.shape[0] != 10 or img.shape[1] != 6:
        img = cv2.resize(img, (6, 10), interpolation=cv2.INTER_CUBIC)
    normalized_img = np.zeros((6, 10))  # 归一化
    normalized_img = cv2.normalize(img, normalized_img, 0, 1, cv2.NORM_MINMAX)
    cv2.imwrite(filename, normalized_img)

归一化的类型,可以有以下的取值:

  • NORM_MINMAX:数组的数值被平移或缩放到一个指定的范围,线性归一化,一般较常用。
  • NORM_INF:此类型的定义没有查到,根据OpenCV 1的对应项,可能是归一化数组的C-范数(绝对值的最大值)
  • NORM_L1 :  归一化数组的L1-范数(绝对值的和)
  • NORM_L2: 归一化数组的(欧几里德)L2-范数

字符识别

字符图片 宽6个像素,高10个像素 ,理论上可以最简单粗暴地可以定义出60个特征:60个像素点上面的像素值。但是显然这样高维度必然会造成过大的计算量,可以适当的降维。比如:

  • 每行上黑色像素的个数,可以得到10个特征
  • 每列上黑色像素的个数,可以得到6个特征
from sklearn.neighbors import KNeighborsClassifier
import os
from sklearn import preprocessing
import cv2
import numpy as np
import warnings
warnings.filterwarnings(module='sklearn*', action='ignore', category=DeprecationWarning)
 
 
def get_feature(file_name):
    img = cv2.imread(file_name, cv2.IMREAD_GRAYSCALE)  # 读取图片时采用单通道
    height, width = img.shape
 
    pixel_cnt_list = []
    for y in range(height):
        pix_cnt_x = 0
        for x in range(width):
            if img[y, x] == 0:  # 黑色点
                pix_cnt_x += 1
 
        pixel_cnt_list.append(pix_cnt_x)
 
    for x in range(width):
        pix_cnt_y = 0
        for y in range(height):
            if img[y, x] == 0:  # 黑色点
                pix_cnt_y += 1
 
        pixel_cnt_list.append(pix_cnt_y)
 
    return pixel_cnt_list
 
 
if __name__ == "__main__":
    test = get_feature("dataset/K/04a0844c-12f2-4344-9b78-ac1d28d746c0.png")
    category = []
    features = []
    for dir_name in os.listdir('dataset'):
        for filename in os.listdir('dataset/' + dir_name):
            category.append(dir_name)
            current_file = 'dataset/' + dir_name + '/' + filename
            feature = get_feature(current_file)
            features.append(feature)
            # print(current_file)
    le = preprocessing.LabelEncoder()
    label = le.fit_transform(category)
 
    model = KNeighborsClassifier(n_neighbors=1)
    model.fit(features, label)
    predicted= model.predict(np.array(test).reshape(1, -1))
    print(predicted)
    print(le.inverse_transform(predicted))

这里直接使用了sklearn中的KNN方法,如需了解更多见:使用 Scikit-learn 进行 KNN 分类

打赏作者 微信支付
《KNN算法实战:验证码的识别》 支付宝
《KNN算法实战:验证码的识别》

    原文作者:算法小白
    原文地址: https://juejin.im/entry/5c66a2dc5188256ec63ef6b5
    本文转自网络文章,转载此文章仅为分享知识,如有侵权,请联系博主进行删除。
点赞