def find_template(im_source,
im_search,
threshold: float = 0.85,
mode=cv2.TM_CCOEFF_NORMED):
"""
模板匹配
:param im_source: 待匹配图像
:param im_search: 待匹配模板
:param threshold: 匹配度
:param mode: 识别模式
:return: None or Rect
"""
start = time.time()
im_source, im_search = check_detection_input(im_source, im_search)
# 模板匹配取得res矩阵
res = _get_template_result_matrix(im_source, im_search)
# 找到最佳匹配项
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
h, w = im_search.shape[:2]
# 求可信度
img_crop = im_source[max_loc[1]:max_loc[1] + h, max_loc[0]:max_loc[0] + w]
confidence = cal_rgb_confidence(img_crop, im_search)
# 如果可信度小于threshold,则返回None
if confidence < threshold:
return None
# 求取位置
x, y = max_loc
rect = Rect(x=x, y=y, width=w, height=h)
print('[tpl]{Rect}, confidence={confidence}, time={time:.2f}'.format(
confidence=confidence, Rect=rect, time=(time.time() - start) * 1000))
return generate_result(rect, confidence)
def find_template(self, im_source, im_search, threshold=None):
"""
模板匹配, 返回匹配度最高的坐标
:param im_source: 待匹配图像
:param im_search: 待匹配模板
:param threshold: 匹配度
:return: None or Rect
"""
start = time.time()
im_source, im_search = self.check_detection_input(im_source, im_search)
# 模板匹配取得res矩阵
res = self._get_template_result_matrix(im_source, im_search)
# 找到最佳匹配项
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
h, w = im_search.shape[:2]
# 求可信度
img_crop = im_search.adjustROI(max_loc[1], max_loc[1] + h, max_loc[0],
max_loc[0] + w)
confidence = self.cal_rgb_confidence(img_crop, im_search)
# confidence = self.test_cal_rgb_confidence(img_crop, im_search)
# 如果可信度小于threshold,则返回None
if confidence < (threshold or self.threshold):
return None
# 求取位置
x, y = max_loc
rect = Rect(x=x, y=y, width=w, height=h)
print('[tpl]{Rect}, confidence={confidence}, time={time:.2f}'.format(
confidence=confidence,
Rect=rect,
time=(time.time() - start) * 1000))
return generate_result(rect, confidence)
def crop_image(self, rect):
"""区域范围截图"""
img = self.imread()
height, width = self.shape
if isinstance(rect, (list, tuple)) and len(rect) == 4:
rect = Rect(*rect)
elif isinstance(rect, Rect):
pass
else:
raise ValueError('unknown rect: type={}, rect={}'.format(type(rect), rect))
if not Rect(0, 0, width, height).contains(rect):
raise OverflowError('Rect不能超出屏幕 {}'.format(rect))
# 获取在图像中的实际有效区域：
x_min, y_min = int(rect.tl.x), int(rect.tl.y)
x_max, y_max = int(rect.br.x), int(rect.br.y)
return image(img[y_min:y_max, x_min:x_max])
def create_similar_rect(x, y, w, h):
x = [x, x + 1, x - 1]
y = [y, y + 1, y - 1]
w = [w, w + 1, w - 1]
h = [h, h + 1, h - 1]
t = []
for i in itertools.product(*[range(3) for k in range(4)]):
if (i[2] == 2 and i[3] == 1) or (i[2] == 1 and i[3] == 2):
pass
else:
t.append(Rect(x=x[i[0]], y=y[i[1]], width=w[i[2]], height=h[i[3]]))
return t
def _two_good_points(self, pts_sch1, pts_sch2, pts_src1, pts_src2,
im_search, im_source):
"""返回两对匹配特征点情形下的识别结果."""
# 先算出中心点(在im_source中的坐标)：
middle_point = [
int((pts_src1[0] + pts_src2[0]) / 2),
int((pts_src1[1] + pts_src2[1]) / 2)
]
pypts = []
# 如果特征点同x轴或同y轴(无论src还是sch中)，均不能计算出目标矩形区域来，此时返回值同good=1情形
if pts_sch1[0] == pts_sch2[0] or pts_sch1[1] == pts_sch2[1] or pts_src1[0] == pts_src2[0] or pts_src1[1] == \
pts_src2[1]:
confidence = self.ONE_POINT_CONFI
return dict(result=middle_point,
rectangle=pypts,
confidence=confidence)
# 计算x,y轴的缩放比例：x_scale、y_scale，从middle点扩张出目标区域:(注意整数计算要转成浮点数结果!)
h, w = im_search.shape[:2]
h_s, w_s = im_source.shape[:2]
x_scale = abs(1.0 * (pts_src2[0] - pts_src1[0]) /
(pts_sch2[0] - pts_sch1[0]))
y_scale = abs(1.0 * (pts_src2[1] - pts_src1[1]) /
(pts_sch2[1] - pts_sch1[1]))
# 得到scale后需要对middle_point进行校正，并非特征点中点，而是映射矩阵的中点。
sch_middle_point = int((pts_sch1[0] + pts_sch2[0]) / 2), int(
(pts_sch1[1] + pts_sch2[1]) / 2)
middle_point[0] = middle_point[0] - int(
(sch_middle_point[0] - w / 2) * x_scale)
middle_point[1] = middle_point[1] - int(
(sch_middle_point[1] - h / 2) * y_scale)
middle_point[0] = max(middle_point[0], 0) # 超出左边界取0 (图像左上角坐标为0,0)
middle_point[0] = min(middle_point[0], w_s - 1) # 超出右边界取w_s-1
middle_point[1] = max(middle_point[1], 0) # 超出上边界取0
middle_point[1] = min(middle_point[1], h_s - 1) # 超出下边界取h_s-1
# 计算出来rectangle角点的顺序：左上角->左下角->右下角->右上角， 注意：暂不考虑图片转动
# 超出左边界取0, 超出右边界取w_s-1, 超出下边界取0, 超出上边界取h_s-1
x_min, x_max = int(max(middle_point[0] - (w * x_scale) / 2, 0)), int(
min(middle_point[0] + (w * x_scale) / 2, w_s - 1))
y_min, y_max = int(max(middle_point[1] - (h * y_scale) / 2, 0)), int(
min(middle_point[1] + (h * y_scale) / 2, h_s - 1))
# 目标矩形的角点按左上、左下、右下、右上的点序：(x_min,y_min)(x_min,y_max)(x_max,y_max)(x_max,y_min)
pts = np.float32([[x_min, y_min], [x_min, y_max], [x_max, y_max],
[x_max, y_min]]).reshape(-1, 1, 2)
for npt in pts.astype(int).tolist():
pypts.append(tuple(npt[0]))
return Rect(x=x_min,
y=y_min,
width=(x_max - x_min),
height=(y_max - y_min))
def find_templates(im_source,
im_search,
threshold: float = 0.85,
max_count=10):
"""
模板匹配
:param im_source: 待匹配图像
:param im_search: 待匹配模板
:param threshold: 匹配度
:param mode: 识别模式
:param max_count: 最多匹配数量
:return: None or Rect
"""
start = time.time()
im_source, im_search = check_detection_input(im_source, im_search)
# 模板匹配取得res矩阵
res = _get_template_result_matrix(im_source, im_search)
result = []
h, w = im_search.shape[:2]
while True:
min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(res)
img_crop = im_source[max_loc[1]:max_loc[1] + h,
max_loc[0]:max_loc[0] + w]
confidence = cal_rgb_confidence(img_crop, im_search)
if confidence < threshold or len(result) > max_count:
break
x, y = max_loc
rect = Rect(x=x, y=y, width=w, height=h)
result.append(generate_result(rect, confidence))
# 屏蔽最优结
cv2.rectangle(res, (int(max_loc[0] - w / 2), int(max_loc[1] - h / 2)),
(int(max_loc[0] + w / 2), int(max_loc[1] + h / 2)),
(0, 0, 0), -1)
if result:
print('[tpls] find counts:{counts}, time={time:.2f}ms{result}'.format(
counts=len(result),
time=(time.time() - start) * 1000,
result=''.join([
'\n\t{}, confidence={}'.format(x['rect'], x['confidence'])
for x in result
])))
return result if result else None
def _many_good_pts(self, im_source, im_search, kp_sch, kp_src,
good) -> Rect:
sch_pts, img_pts = np.float32([
kp_sch[m.queryIdx].pt for m in good
]).reshape(-1, 1, 2), np.float32([kp_src[m.trainIdx].pt
for m in good]).reshape(-1, 1, 2)
M, mask = self._find_homography(sch_pts, img_pts)
# 计算四个角矩阵变换后的坐标，也就是在大图中的目标区域的顶点坐标:
h, w = im_search.shape[:2]
h_s, w_s = im_source.shape[:2]
pts = np.float32([[0, 0], [0, h - 1], [w - 1, h - 1],
[w - 1, 0]]).reshape(-1, 1, 2)
dst = cv2.perspectiveTransform(pts, M)
# trans numpy arrary to python list: [(a, b), (a1, b1), ...]
def cal_rect_pts(dst):
return [
tuple(npt[0])
for npt in np.rint(dst).astype(np.float).tolist()
]
pypts = cal_rect_pts(dst)
# pypts四个值按照顺序分别是: 左上,左下,右下,右上
# 注意：虽然4个角点有可能越出source图边界，但是(根据精确化映射单映射矩阵M线性机制)中点不会越出边界
lt, br = pypts[0], pypts[2]
# 考虑到算出的目标矩阵有可能是翻转的情况，必须进行一次处理，确保映射后的“左上角”在图片中也是左上角点：
x_min, x_max = min(lt[0], br[0]), max(lt[0], br[0])
y_min, y_max = min(lt[1], br[1]), max(lt[1], br[1])
# 挑选出目标矩形区域可能会有越界情况，越界时直接将其置为边界：
# 超出左边界取0，超出右边界取w_s-1，超出下边界取0，超出上边界取h_s-1
# 当x_min小于0时，取0。 x_max小于0时，取0。
x_min, x_max = int(max(x_min, 0)), int(max(x_max, 0))
# 当x_min大于w_s时，取值w_s-1。 x_max大于w_s-1时，取w_s-1。
x_min, x_max = int(min(x_min, w_s - 1)), int(min(x_max, w_s - 1))
# 当y_min小于0时，取0。 y_max小于0时，取0。
y_min, y_max = int(max(y_min, 0)), int(max(y_max, 0))
# 当y_min大于h_s时，取值h_s-1。 y_max大于h_s-1时，取h_s-1。
y_min, y_max = int(min(y_min, h_s - 1)), int(min(y_max, h_s - 1))
return Rect(x=x_min,
y=y_min,
width=(x_max - x_min),
height=(y_max - y_min))
from core.cv.match_template import find_template
Anchor = Anchor(dev={
'width': 1920,
'height': 1080
},
cur={
'width': 3400,
'height': 1440,
'left': 260,
'right': 260
},
orientation=1)
rect = Rect.create_by_point_size(Anchor.point(0, 0), Anchor.size(1920, 1080))
img = image('./core/cv/test_image/emulator-5554.png')
#
im_search = image('./core/cv/test_image/star.png').resize(
62 * 1.33333, 43 * 1.33333)
#
#
# a = cuda_find_template(im_source=img, im_search=im_search)
import cv2
from core.cv.cuda_match_template import cuda_template
tpl = cuda_template()
for i in range(10):
tpl.find_template(im_source=img, im_search=img)
# img2 = cv2.cuda_GpuMat()
# img2.upload(cv2.imread('./core/cv/test_image/star.png'))