《百万英雄》Python + OCR搜索引擎统计辅助简单实现思路

0. 环境

iOS 10 + MacOS 10.12 + Python 2.7

1. 思路

APP界面中弹出的题的位置和答案的位置都是固定的,因此我们可以将手机屏幕想办法投到电脑屏幕上,通过OCR识别指定区域,实时打开搜索引擎界面搜索问题,甚至匹配答案。

2. 关键步骤

2.1 投屏

我是iPhone 5s + Mac电脑,可以用Mac的Quicktime Player播放器的屏幕录制功能(安卓据说可以用ADB)。

具体的,打开Quicktime Player后,点击“文件–新建屏幕录制–(红色录制按钮旁的下拉菜单)选择从手机录制”,这时,手机屏幕就实时投到屏幕上了。

2.2 截屏和OCR

  • 截屏

截屏要将你的手机投屏窗口固定在一个位置,找准屏幕上的左上角和右下角两个坐标,利用PIL中的ImageGrab进行抓屏,以截取题干为例,代码如下:

from PIL import ImageGrab
image = ImageGrab.grab((50, 170, 540, 330))
  • OCR

利用tesseract库和对应的pytesseract接口进行OCR,具体配置可以参考[1]。

import pytesseract
ocr_str = pytesseract.image_to_string(image, lang='chi_sim')

2.3 搜索

分两种思路,我们可以直接打开一个浏览器页面用百度搜索,把答案筛选工作交给人:

import webbrowser
url = "http://www.baidu.com/s?rn=50&wd=" + ocr_str.encode(encoding='UTF-8',errors='strict')
webbrowser.open_new_tab(url)

也可以将搜索结果页面下载下来用选项字符串匹配,统计该出现的次数(当然,第二种方法需要增加一次识别答案字符串的OCR过程):

# 以统计答案1出现的次数为例
import urllib
res = urllib.urlopen(url).read()
o1cnt = res.count(o1_ocr.encode(encoding='UTF-8'))

3. 优化

3.1 分词

尤其是在我们用选项字符串匹配下载下来的搜索页面文本时,很可能匹配数很少,这是由于正确答案不一定一字不差地藏在搜索文本中,我想到的更好的方法就是进行分词,然后匹配出现的次数。这里用到了jieba分词的python接口。这时我们就应该将2.3节中的第二段代码改为如下:

import jieba
import urllib
res = urllib.urlopen(url).read()
o1cnt = 0
for i in o1str_c:
    o1str_c += jieba.cut_for_search(o1str)

3.2 多进程并行

游戏只有10秒钟,而且题干是从左到右滚动出现的,所以留给我们计算的时间只有8秒左右,时间十分重要,利用line profiler工具,发现最耗时的部分出现在OCR部分(2秒左右),如果要进行选项匹配,需要2次截屏和OCR,所以,想到可以用2个线程将两次OCR并行,将选项的OCR放到另一个子线程中,在最后进行字符串匹配时进行同步。

def options(q):
    o = ImageGrab.grab((60, 395, 380, 640))
    ostr = pytesseract.image_to_string(o, lang='chi_sim').encode(encoding='UTF-8',errors='strict')
    ostr_l = ostr.split('\n')
    q.put(ostr_l)

def main():

    # ... 其他初始化

    q = Queue(maxsize = 10)
    o_p = Process(target = options, args = (q, ))
    o_p.start()

    # ... 识别题干的OCR和下载搜索结果页面

    o_p.join()
    ostr_l = q.get()

    # ... 进行选项字符串和搜索结果页面字符串的匹配统计

另外,如果我们还同时打开浏览器页面用于肉眼搜索,打开浏览器也是挺耗时的(0.7秒左右),我们可以将其放到一个子进程中进行。

至此,我们将原先7秒左右可以运行完的程序,优化到了4秒左右,还能留下3秒钟供我们考虑到底选哪个。

实现代码,仅供参考:


#!/usr/local/bin/python
# -*- coding: utf-8 -*-
import pytesseract
from PIL import Image
from PIL import ImageGrab
import webbrowser
import time
import jieba
import urllib
import threading
from multiprocessing import Process, Queue
#DEBUG = True
DEBUG = False
#CUT = False
CUT = True
def start_browser(s):
#pass
webbrowser.open_new_tab(s)
def options(q):
o = ImageGrab.grab((60, 395, 380, 640))
if DEBUG:
o.save('/Users/Jaycee/test/ocr/iphone/options.png')
ostr = pytesseract.image_to_string(o, lang='chi_sim').encode(encoding='UTF-8',errors='strict')
ostr_l = ostr.split('\n')
q.put(ostr_l)
#@profile
def main():
while True:
t00 = time.time()
q = Queue(maxsize = 10)
o_p = Process(target = options, args = (q, ))
o_p.start()
# (y1, x1, y2, x2)
# 50 170 520 320
# 40 140 445 300
image = ImageGrab.grab((50, 170, 540, 330))
t0 = time.time()
#image = ImageGrab.grab((40, 140, 455, 300))
#image.save('/Users/Jaycee/test/ocr/iphone/1.png')
if DEBUG:
t1 = time.time() # grab time
grab_time = t1 – t0
image.save('/Users/Jaycee/test/ocr/iphone/1.png')
#image = Image.open('/Users/Jaycee/test/ocr/iphone/1.png')
t1 = time.time()
# open image
#code = pytesseract.image_to_string(image, lang='chi_sim').encode(encoding='UTF-8',errors='strict')
code = pytesseract.image_to_string(image, lang='chi_sim')
print code
if CUT:
jieba_s = jieba.cut_for_search(code)
jieba_s = ' '.join(jieba_s)
code = jieba_s
if DEBUG:
t2 = time.time() # ocr time
ocr_time = t2 – t1
url = "http://www.baidu.com/s?rn=50&wd=" + code.encode(encoding='UTF-8',errors='strict')
#url = "https://www.google.com/search?q=" + code.encode(encoding='UTF-8',errors='strict')
p = Process(target = start_browser, args = (url, ))
p.start()
if DEBUG:
t3 = time.time()
open_browser_time = t3 – t2
t000 = time.time()
res = urllib.urlopen(url).read()
t111 = time.time()
print "Download Html Time:", t111 – t000
o_p.join()
ostr_l = q.get()
try:
o1str = ostr_l[0]
o2str = ostr_l[2]
o3str = ostr_l[4]
except:
a = raw_input("Error! Press 'Enter' to process next..")
continue
else:
pass
o1str_c = jieba.cut_for_search(o1str)
o2str_c = jieba.cut_for_search(o2str)
o3str_c = jieba.cut_for_search(o3str)
o1cnt = 0
o2cnt = 0
o3cnt = 0
for i in o1str_c:
o1cnt += res.count(i.encode(encoding='UTF-8'))
for i in o2str_c:
o2cnt += res.count(i.encode(encoding='UTF-8'))
for i in o3str_c:
o3cnt += res.count(i.encode(encoding='UTF-8'))
if DEBUG:
t4 = time.time()
option_count_time = t4 – t3
print "grab time:", grab_time, "ocr time:", ocr_time, "open browser time:", open_browser_time, "open_browser_time:", open_browser_time
print "A:\t[OCR] %s [COUNT] %d" % (o1str, o1cnt)
print "B:\t[OCR] %s [COUNT] %d" % (o2str, o2cnt)
print "C:\t[OCR] %s [COUNT] %d" % (o3str, o3cnt)
t11 = time.time()
print "Total Time:", t11-t00
a = raw_input("Press 'Enter' to process next..")
if __name__ == "__main__":
main()

view raw

bwfw.py

hosted with ❤ by GitHub


[1] Python 中文OCR, http://blog.csdn.net/wwj_748/article/details/78109680?utm_source=tuicool&utm_medium=referral

2 comments

Leave a Reply to JayceeZ Cancel reply

Your email address will not be published. Required fields are marked *