斗地主AI出牌助手--在线调用斗地主AI，实现自动斗地主

简介

程序基于Python3.7开发的斗地主AI出牌助手，目前支持欢乐斗地主桌面版，微信版，也可以自己制作相应其他版本。

此出牌助手核心是识别出三位玩家出牌内容，调用基于DouZero封装的API接口，输入出牌内容，根据AI出牌方案，打出相应的牌。

运行效果

![](https://img2023.cnblogs.com/blog/129408/202311/129408-20231114235431212-40412513.gif)

核心功能

• 手牌和位置识别
  1. 游戏刚开始根据屏幕位置，截图识别AI玩家手牌及三张底牌
  2. 根据玩家手牌判断是否抢地主和加倍
  3. 确认三者之间的关系，识别地主和农民角色，确认队友及对手关系
• 识别每轮三位玩家出牌
  1. 根据提示按钮，判断当前的操作
  2. 识别三位玩家此轮出的牌，并记录下来
• AI出牌方案输出
  1. 将出牌记录按格式要求发送给斗地主AI
  2. 获取斗地主AI出牌方案，选取最高胜率方案
  3. 根据AI出牌方案，选择对相应的牌，并出牌

素材准备

• 区域定位，获取坐标值
  
  

• 按钮及牌面
  
  

核心代码

• 根据指定窗口句柄截图

  def WindowShot(self):
      """
      根据窗口句柄截图
      返回: 图片对象
      """
      windll.user32.SetProcessDPIAware()
      hwnd = self.Handle
      left, top, right, bottom = win32gui.GetClientRect(hwnd)
      w = right - left
      h = bottom - top
      hwnd_dc = win32gui.GetWindowDC(hwnd)
      mfc_dc = win32ui.CreateDCFromHandle(hwnd_dc)
      save_dc = mfc_dc.CreateCompatibleDC()
      bitmap = win32ui.CreateBitmap()
      bitmap.CreateCompatibleBitmap(mfc_dc, w, h)
      save_dc.SelectObject(bitmap)
      # If Special K is running, this number is 3. If not, 1
      result = windll.user32.PrintWindow(hwnd, save_dc.GetSafeHdc(), 3)
      bmpinfo = bitmap.GetInfo()
      bmpstr = bitmap.GetBitmapBits(True)
      img = np.frombuffer(bmpstr, dtype=np.uint8).reshape((bmpinfo["bmHeight"], bmpinfo["bmWidth"], 4))
      img = np.ascontiguousarray(img)[..., :-1]  # make image C_CONTIGUOUS and drop alpha channel
      #img = Image.frombuffer("RGB",(bmpinfo['bmWidth'], bmpinfo['bmHeight']),bmpstr, 'raw', 'BGRX', 0, 1)
      if not result:  # result should be 1
          win32gui.DeleteObject(bitmap.GetHandle())
          save_dc.DeleteDC()
          mfc_dc.DeleteDC()
          win32gui.ReleaseDC(hwnd, hwnd_dc)
          raise RuntimeError(f"Unable to acquire screenshot! Result: {result}")
      #cv2.imwrite('./imgs/print.png', img)
      #return img
      return cv2.cvtColor(img, cv2.COLOR_BGRA2BGR)
  

• 在图片中查找第一个相似的图片

  def LocateOnImage(self, template, image=None, region=None, confidence=0.9, grayscale=True):
      """
      在image中寻找template，返回第一个查找到的范围
  
      参数:
      template: 需要查找的图片，文件名或图片对象
      image: 被查找的图片，文件名或图片对象
      region: 查找范围
      confidence: 置信度
      grayscale: 是否为灰度图
  
      返回值:
      查找到的图片范围
      """
      if image is None:
          image = self.WindowShot()
      return pyautogui.locate(template, image, region=region, confidence=confidence, grayscale=grayscale)
  

• 在图片中查找所有相似的图片

  def LocateAllOnImage(self, template, image=None, region=None, confidence=0.9, grayscale=True):
      """
      在image中寻找template，返回第一个查找到的范围
  
      参数:
      template: 需要查找的图片，文件名或图片对象
      image: 被查找的图片，文件名或图片对象
      region: 查找范围
      confidence: 置信度
      grayscale: 是否为灰度图
  
      返回值:
      查找到的图片范围
      """
      if image is None:
          image = self.WindowShot()
      return pyautogui.locateAll(template, image, region=region, confidence=confidence, grayscale=grayscale)
  

• 识别手牌

  def GetCards(self, image, player):
      hand_cards = []
      cards = ""
      start_x = 0
      width = player["width"]
      for card in self.AllCardsNC:
          confidence = player["confidence"] or 0.85
          grayscale = True
          if(card in ('D', 'X')):
              confidence = 0.85
              grayscale = False
  
          card_key = player["prefix"] + card
          matches = self.LocateAllOnImageName(card_key, image, player["region"], confidence, grayscale)
  
          if len(matches) > 0:
              sorted_matches = sorted(matches, key=lambda match: match[0])
              #print(target_position)
              #print(mark)
  
              if(card != 'X'):
                  start_x = 0
  
              for match in sorted_matches:
                  if(start_x == 0 or start_x + player["width"] < match[0]):
                      #大小王容易判断错误，需要再判断一下颜色值
                      if(card == 'D'):
                          #match = sorted_matches[0]
                          y,x,h,w = match
                          cropped_image = image[x:x+w, y:y+h]
                          #colors = cv2.mean(cropped_image)
                          mean, stddev = cv2.meanStdDev(cropped_image)
                          #print("均值：", mean)
                          #print("差值：", stddev)
                          if(stddev[2][0] > 30):
                              continue
                      start_x = match[0]
                      width = match[2]
                      hand_cards.append({card:match})
                      cards += card
  
      return cards, hand_cards
  

• 将出牌记录发送给AI，进行预测

  def GetCardsForPredict(self):
      #将出牌记录发送给AI，进行预测
      resultStr = self.PostPredict()
      #获取最高胜率的出牌
      last_move_cards = self.GetPredictWinRates(resultStr)
      return last_move_cards
  

• 出牌

  def PlayedCards(self, cards, image=None):
      if image is not None:
          window_shot_image = image
      else:
          window_shot_image = self.WindowShot()
      player_self_cards, range = self.GetCards(window_shot_image, self.Players["PlayerSelf"])
      select_index = []
      select_cards = []
      #将cards中的字符顺序反转
      play_cards = cards[::-1]
      cards = play_cards
  
      for c in cards:
          index = -1
          for card in player_self_cards:
              index += 1
  
              if c == card and index not in select_index:
                  select_index.append(index)
                  select_cards.append(c)
                  #print(range[index])
                  cards = cards.replace(c, '', 1)
                  card_range = range[index][c]
                  break
  
      cards = play_cards
      index = -1
      #window_shot_image = self.WindowShot()
      for n in select_index:
          window_shot_image = self.WindowShot()
          player_self_cards, range = self.GetCards(window_shot_image, self.Players["PlayerSelf"])
          if(len(player_self_cards) > n):
              #player_self_cards = player_self_cards
              index += 1
              card_range = range[n][cards[index]]
              #如果要出的牌还未选择完毕，则点击
              if(card_range[1] > self.CardTop):
                  self.LeftClick(card_range)
  
          time.sleep(0.2)
  
      return self.GetSelectCards(play_cards)
  

• DouZero 的请求结构体

  Predict = {
      "bomb_num":0,#炸弹数量
      "card_play_action_seq":'',#历史出牌动作序列，用逗号分隔
      "last_move_landlord":'',#地主最后出的牌
      "last_move_landlord_down":'',#地主下家最后出的牌
      "last_move_landlord_up":'',#地主上家最后出的牌
      "num_cards_left_landlord":20,#地主手牌剩余数量
      "num_cards_left_landlord_down":17,#地主下家手牌剩余数量
      "num_cards_left_landlord_up":17,#地主上家手牌剩余数量
      "other_hand_cards":'',#还剩余的牌
      "played_cards_landlord":'',#地主所有出的牌
      "played_cards_landlord_down":'',#地主下家所有出的牌
      "played_cards_landlord_up":'',#地主上家所有出的牌
      "player_hand_cards":'',#玩家手中的牌
      "player_position":0,	#-当前玩家的位置序号 0 地主，1 地主下家，2 地主上家
      "three_landlord_cards":''#三张底牌
  }
  

可用的 DouZero 后端地址

• 快手的 DouZero
  
  https://douzero.org:5000/predict
• 网易的 PerfectDou
  
  https://outer-perfectdou-demo-gzailab.nie.netease.com:10074/predict
• 也可根据 RLCard-Showdown 自行搭建本地运行的后端服务

免责声明

本程序仅供娱乐和学习使用，不得用于任何非法用途

参考项目

• DouZero：https://github.com/kwai/DouZero
• PerfectDou：https://github.com/Netease-Games-AI-Lab-Guangzhou/PerfectDou
• rlcard-showdown：https://github.com/datamllab/rlcard-showdown
• DouZero_For_HLDDZ_FullAuto: https://github.com/Vincentzyx/DouZero_For_HLDDZ_FullAuto