# !/usr/bin/env python
# -*- coding: utf-8 -*-
"""多尺度模板匹配.
对用户提供的调节参数:
1. threshod: 筛选阈值，默认为0.8
2. rgb: 彩色三通道,进行彩色权识别.
3. scale_max: 多尺度模板匹配最大范围，增大可适应更小UI
4. scale_step: 多尺度模板匹配搜索比例步长，减小可适应更小UI
"""
from __future__ import division
from __future__ import print_function
import time
import cv2
from ha4t.aircv.aircv import crop_image
from ha4t.aircv.cal_confidence import cal_rgb_confidence, cal_ccoeff_confidence
from ha4t.aircv.error import TemplateInputError
from ha4t.aircv.utils import generate_result, check_source_larger_than_search, img_mat_rgb_2_gray, print_run_time
[文档]
class MultiScaleTemplateMatching(object):
"""多尺度模板匹配."""
METHOD_NAME = "MSTemplate"
def __init__(self, im_search, im_source, threshold=0.8, rgb=True, record_pos=None, resolution=(), scale_max=800, scale_step=0.005):
self.im_source = im_source
self.im_search = im_search
self.threshold = threshold
self.rgb = rgb
self.record_pos = record_pos
self.resolution = resolution
self.scale_max = scale_max
self.scale_step = scale_step
[文档]
def find_all_results(self):
raise NotImplementedError
@print_run_time
def find_best_result(self):
"""函数功能：找到最优结果."""
# 第一步：校验图像输入
check_source_larger_than_search(self.im_source, self.im_search)
# 第二步：计算模板匹配的结果矩阵res
s_gray, i_gray = img_mat_rgb_2_gray(
self.im_search), img_mat_rgb_2_gray(self.im_source)
confidence, max_loc, w, h, _ = self.multi_scale_search(
i_gray, s_gray, ratio_min=0.01, ratio_max=0.99, src_max=self.scale_max, step=self.scale_step, threshold=self.threshold)
# 求取识别位置: 目标中心 + 目标区域:
middle_point, rectangle = self._get_target_rectangle(max_loc, w, h)
best_match = generate_result(middle_point, rectangle, confidence)
# LOGGING.debug("[%s] threshold=%s, result=%s" %
# (self.METHOD_NAME, self.threshold, best_match))
return best_match if confidence >= self.threshold else None
def _get_confidence_from_matrix(self, max_loc, w, h):
"""根据结果矩阵求出confidence."""
sch_h, sch_w = self.im_search.shape[0], self.im_search.shape[1]
if self.rgb:
# 如果有颜色校验,对目标区域进行BGR三通道校验:
img_crop = self.im_source[max_loc[1]:max_loc[1] + h, max_loc[0]: max_loc[0] + w]
confidence = cal_rgb_confidence(
cv2.resize(img_crop, (sch_w, sch_h)), self.im_search)
else:
img_crop = self.im_source[max_loc[1]:max_loc[1] + h, max_loc[0]: max_loc[0] + w]
confidence = cal_ccoeff_confidence(
cv2.resize(img_crop, (sch_w, sch_h)), self.im_search)
return confidence
def _get_target_rectangle(self, left_top_pos, w, h):
"""根据左上角点和宽高求出目标区域."""
x_min, y_min = left_top_pos
# 中心位置的坐标:
x_middle, y_middle = int(x_min + w / 2), int(y_min + h / 2)
# 左下(min,max)->右下(max,max)->右上(max,min)
left_bottom_pos, right_bottom_pos = (
x_min, y_min + h), (x_min + w, y_min + h)
right_top_pos = (x_min + w, y_min)
# 点击位置:
middle_point = (x_middle, y_middle)
# 识别目标区域: 点序:左上->左下->右下->右上, 左上(min,min)右下(max,max)
rectangle = (left_top_pos, left_bottom_pos,
right_bottom_pos, right_top_pos)
return middle_point, rectangle
@staticmethod
def _resize_by_ratio(src, templ, ratio=1.0, templ_min=10, src_max=800):
"""根据模板相对屏幕的长边 按比例缩放屏幕"""
# 截屏最大尺寸限制
sr = min(src_max/max(src.shape),1.0)
src = cv2.resize(src, (int(src.shape[1]*sr), int(src.shape[0]*sr)))
# 截图尺寸缩放
h, w = src.shape[0], src.shape[1]
th, tw = templ.shape[0], templ.shape[1]
if th/h >= tw/w:
tr = (h*ratio)/th
else:
tr = (w*ratio)/tw
templ = cv2.resize(templ, (max(int(tw*tr), 1), max(int(th*tr), 1)))
return src, templ, tr, sr
@staticmethod
def _org_size(max_loc, w, h, tr, sr):
"""获取原始比例的框"""
max_loc = (int((max_loc[0]/sr)), int((max_loc[1]/sr)))
w, h = int((w/sr)), int((h/sr))
return max_loc, w, h
[文档]
def multi_scale_search(self, org_src, org_templ, templ_min=10, src_max=800, ratio_min=0.01,
ratio_max=0.99, step=0.01, threshold=0.8, time_out=3.0):
"""多尺度模板匹配"""
mmax_val = 0
max_info = None
r = ratio_min
t = time.time()
while r <= ratio_max:
src, templ, tr, sr = self._resize_by_ratio(
org_src.copy(), org_templ.copy(), r, src_max=src_max)
if min(templ.shape) > templ_min:
src[0,0] = templ[0,0] = 0
src[0,1] = templ[0,1] = 255
result = cv2.matchTemplate(src, templ, cv2.TM_CCOEFF_NORMED)
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(result)
h, w = templ.shape
if mmax_val < max_val:
mmax_val = max_val
max_info = (r, max_val, max_loc, w, h, tr, sr)
# print((r, max_val, max_loc, w, h, tr, sr))
time_cost = time.time() - t
if time_cost>time_out and max_val >= threshold:
omax_loc, ow, oh = self._org_size(max_loc, w, h, tr, sr)
confidence = self._get_confidence_from_matrix(omax_loc, ow, oh)
if confidence >= threshold:
return confidence, omax_loc, ow, oh, r
r += step
if max_info is None:
return 0, (0, 0), 0, 0, 0
max_r, max_val, max_loc, w, h, tr, sr = max_info
omax_loc, ow, oh = self._org_size(max_loc, w, h, tr, sr)
confidence = self._get_confidence_from_matrix(omax_loc, ow, oh)
return confidence, omax_loc, ow, oh, max_r
[文档]
class MultiScaleTemplateMatchingPre(MultiScaleTemplateMatching):
"""基于截图预设条件的多尺度模板匹配."""
METHOD_NAME = "MSTemplatePre"
DEVIATION = 150
@print_run_time
def find_best_result(self):
"""函数功能：找到最优结果."""
if self.resolution!=():
# 第一步：校验图像输入
check_source_larger_than_search(self.im_source, self.im_search)
if self.resolution[0]<self.im_search.shape[1] or self.resolution[1]<self.im_search.shape[0]:
raise TemplateInputError("error: resolution is too small.")
# 第二步：计算模板匹配的结果矩阵res
if not self.record_pos is None:
area, self.resolution = self._get_area_scope(self.im_source, self.im_search, self.record_pos, self.resolution)
self.im_source = crop_image(self.im_source, area)
check_source_larger_than_search(self.im_source, self.im_search)
r_min, r_max = self._get_ratio_scope(
self.im_source, self.im_search, self.resolution)
s_gray, i_gray = img_mat_rgb_2_gray(self.im_search), img_mat_rgb_2_gray(self.im_source)
confidence, max_loc, w, h, _ = self.multi_scale_search(
i_gray, s_gray, ratio_min=r_min, ratio_max=r_max, step=self.scale_step,
threshold=self.threshold, time_out=1.0)
if not self.record_pos is None:
max_loc = (max_loc[0] + area[0], max_loc[1] + area[1])
# 求取识别位置: 目标中心 + 目标区域:
middle_point, rectangle = self._get_target_rectangle(max_loc, w, h)
best_match = generate_result(middle_point, rectangle, confidence)
# LOGGING.debug("[%s] threshold=%s, result=%s" %
# (self.METHOD_NAME, self.threshold, best_match))
return best_match if confidence >= self.threshold else None
else:
return None
def _get_ratio_scope(self, src, templ, resolution):
"""预测缩放比的上下限."""
H, W = src.shape[0], src.shape[1]
th, tw = templ.shape[0], templ.shape[1]
w, h = resolution
rmin = min(H/h, W/w) # 新旧模板比下限
rmax = max(H/h, W/w) # 新旧模板比上限
ratio = max(th/H, tw/W) # 小图大图比
r_min = ratio*rmin
r_max = ratio*rmax
return max(r_min, self.scale_step), min(r_max, 0.99)
[文档]
def get_predict_point(self, record_pos, screen_resolution):
"""预测缩放后的点击位置点."""
delta_x, delta_y = record_pos
_w, _h = screen_resolution
target_x = delta_x * _w + _w * 0.5
target_y = delta_y * _w + _h * 0.5
return target_x, target_y
def _get_area_scope(self, src, templ, record_pos, resolution):
"""预测搜索区域."""
H, W = src.shape[0], src.shape[1]
th, tw = templ.shape[0], templ.shape[1]
w, h = resolution
x, y = self.get_predict_point(record_pos, (W, H))
predict_x_radius = max(int(tw * W / (w)), self.DEVIATION)
predict_y_radius = max(int(th * H / (h)), self.DEVIATION)
area = (
max(x - predict_x_radius, 0),
max(y - predict_y_radius, 0),
min(x + predict_x_radius, W),
min(y + predict_y_radius, H),
)
return area, (w*(area[3]-area[1])/W, h*(area[2]-area[0])/H)