这项功能的目的是为了方便使用opencv做图像标注工具。
在做图像数据标注时，很难一次就做到精准标注，经常需要微调才能达到比较好的标注效果。如果目标比较小，即使微调也难以做到精准，所以就需要另外一个窗口对标注区域进行局部放大以方便微调。
本文中标注信息以矩形框作为示例，矩形框是图像标注中最常用到的一种标注信息形态。其他标注信息的设计逻辑雷同。
程序主要逻辑是：鼠标在任意窗口中做的操作都要同步映射到另外一个窗口。主要通过两个维度进行控制：1. 鼠标在主窗口还是在放大窗口；2. 鼠标的操作。其中主窗口指的是用来显示完整大图的窗口，放大窗口是用来显示局部放大区域的窗口。
总体来说，我们在主窗口做任何操作都可以比较随意，因为主窗口中除了矩形框外，图像本身不发生变化。而在放大窗口中做操作就需要相对保守一些，不然会非常乱，以至于无法操作。所以我们在主窗口中画矩形框时，放大窗口会随着矩形框不停改变；而我们在放大窗口中画矩形框时，放大窗口则保持不变（不然眼花缭乱）。
把程序逻辑搞明白后下面代码就比较容易懂了。细节就不多说了，只说一个需要注意的大方向：我们需要为两个窗口各写一个鼠标回调函数。
先看一下单纯画矩形框的代码可能会有助于理解下面程序中的一些细节：opencv-python鼠标画矩形框：cv2.rectangle()
# -*- coding: utf-8 -*-
import cv2
class Rect(object):
def __init__(self, pt1=(0, 0), pt2=(0, 0)):
self.tl = pt1
self.br = pt2
self.regularize()
def regularize(self):
"""
make sure tl = TopLeft point, br = BottomRight point
"""
tl = (min(self.tl[0], self.br[0]), min(self.tl[1], self.br[1]))
br = (max(self.tl[0], self.br[0]), max(self.tl[1], self.br[1]))
self.tl = tl
self.br = br
def get_center(self):
"""
get center point of Rect
"""
center_x = (self.tl[0] + self.br[0]) // 2
center_y = (self.tl[1] + self.br[1]) // 2
return center_x, center_y
def get_width(self):
"""
get width of Rect
"""
return abs(self.br[0] - self.tl[0])
def get_height(self):
"""
get height of Rect
"""
return abs(self.br[1] - self.tl[1])
def height_over_width(self):
"""
ratio of height over width
"""
return self.get_height() / self.get_width()
def get_area(self):
"""
get area of Rect
"""
return self.get_width() * self.get_height()
class DrawZoom(object):
def __init__(self, image, color,
current_pt=(0, 0),
default_zoom_image_size=(256, 256)):
self.original_image = image
self.color = color
self.thickness = 2
self.current_pt = current_pt
self.default_zoom_image_size = default_zoom_image_size
self.rect_in_big_image = Rect()
self.rect_in_zoom_image = Rect()
self.zoom_offset = (0, 0)
self.is_drawing_big = False
self.is_drawing_zoom = False
self.exist_rect = False
self.big_image = image.copy()
self.zoom_image = None
self.zoom_image_backup = None
self.get_zoom_image()
def get_image_height(self):
"""
get height of big image
"""
return self.original_image.shape[0]
def get_image_width(self):
"""
get width of big image
"""
return self.original_image.shape[1]
def get_margin_height(self):
"""
get height of margin. in the margin area of big image, coordinate of
current_pt does NOT change
"""
return self.default_zoom_image_size[0] // 2
def get_margin_width(self):
"""
get width of margin
"""
return self.default_zoom_image_size[1] // 2
def get_zoom_image(self, height_ratio_expand=0.2, width_ratio_expand=0.2):
"""
get zoom image for two cases: the rect exists or not.
height_ratio_expand and width_ratio_expand are used for expanding some
area of rect
"""
if not self.exist_rect:
self.get_zoom_image_for_current_pt()
elif self.rect_in_big_image.get_area() > 0:
self.get_zoom_image_for_rect(height_ratio_expand,
width_ratio_expand)
def get_zoom_image_for_current_pt(self):
"""
get zoom image for current mouse point (when rect does not exist)
"""
# (x, y) is center coordinate
x = max(self.current_pt[0], self.get_margin_width())
x = min(x, self.get_image_width() - self.get_margin_width())
y = max(self.current_pt[1], self.get_margin_height())
y = min(y, self.get_image_height() - self.get_margin_height())
tl_x = x - self.get_margin_width()
tl_y = y - self.get_margin_height()
br_x = x + self.get_margin_width()
br_y = y + self.get_margin_height()
tl_x, tl_y, br_x, br_y = self.shrink_rect(tl_x, tl_y, br_x, br_y)
self.zoom_image = self.big_image[tl_y:br_y, tl_x:br_x]
self.zoom_image_backup = self.original_image[tl_y:br_y, tl_x:br_x]
self.zoom_offset = (tl_x, tl_y)
def get_zoom_image_for_rect(self, height_ratio_expand, width_ratio_expand):
"""
get zoom image when rect exists
"""
if self.rect_in_big_image.get_area() == 0:
return None
height_over_width_for_win_zoom = \
self.default_zoom_image_size[1] / self.default_zoom_image_size[0]
center = self.rect_in_big_image.get_center()
if self.rect_in_big_image.height_over_width() > \
height_over_width_for_win_zoom:
half_height = int(0.5 * (1 + height_ratio_expand) *
self.rect_in_big_image.get_height())
half_width = int(half_height / height_over_width_for_win_zoom)
else:
half_width = int(0.5 * (1 + width_ratio_expand) *
self.rect_in_big_image.get_width())
half_height = int(half_width * height_over_width_for_win_zoom)
tl_x = center[0] - half_width
tl_y = center[1] - half_height
br_x = center[0] + half_width
br_y = center[1] + half_height
tl_x, tl_y, br_x, br_y = self.shrink_rect(tl_x, tl_y, br_x, br_y)
self.zoom_image = self.big_image[tl_y:br_y, tl_x:br_x]
self.zoom_image_backup = self.original_image[tl_y:br_y, tl_x:br_x]
self.zoom_offset = (tl_x, tl_y)
@staticmethod
def clip(value, low, high):
"""
clip value between low and high
"""
output = max(value, low)
output = min(output, high)
return output
def shrink_point(self, x, y):
"""
shrink point (x, y) to inside big image
"""
x_shrink = self.clip(x, 0, self.get_image_width())
y_shrink = self.clip(y, 0, self.get_image_height())
return x_shrink, y_shrink
def shrink_rect(self, pt1_x, pt1_y, pt2_x, pt2_y):
"""
shrink rect to inside big image
"""
pt1_x = self.clip(pt1_x, 0, self.get_image_width())
pt1_y = self.clip(pt1_y, 0, self.get_image_height())
pt2_x = self.clip(pt2_x, 0, self.get_image_width())
pt2_y = self.clip(pt2_y, 0, self.get_image_height())
rect = Rect((pt1_x, pt1_y), (pt2_x, pt2_y))
rect.regularize()
tl_x, tl_y = rect.tl
br_x, br_y = rect.br
return tl_x, tl_y, br_x, br_y
def reset_big_image(self):
"""
reset big_image (for show) using original image
"""
self.big_image = self.original_image.copy()
def reset_zoom_image(self):
"""
reset zoom_image (for show) using the zoom image backup
"""
self.zoom_image = self.zoom_image_backup.copy()
def draw_rect_in_big_image(self):
"""
draw rect in big image
"""
cv2.rectangle(self.big_image,
self.rect_in_big_image.tl, self.rect_in_big_image.br,
color=self.color, thickness=self.thickness)
def draw_rect_in_zoom_image(self):
"""
draw rect in zoom image
"""
cv2.rectangle(self.zoom_image,
self.rect_in_zoom_image.tl, self.rect_in_zoom_image.br,
color=self.color, thickness=self.thickness)
def update_drawing_big(self):
"""
update drawing big image, and map the corresponding area to zoom image
"""
if self.exist_rect:
self.draw_rect_in_big_image()
self.get_zoom_image()
def update_drawing_zoom(self):
"""
update drawing big and zoom image when drawing rect in zoom image
"""
if self.exist_rect:
self.draw_rect_in_big_image()
self.draw_rect_in_zoom_image()
def onmouse_big_image(event, x, y, flags, draw_zoom):
if event == cv2.EVENT_LBUTTONDOWN:
# pick first point of rect
draw_zoom.is_drawing_big = True
draw_zoom.rect_in_big_image.tl = (x, y)
draw_zoom.exist_rect = True
elif draw_zoom.is_drawing_big and event == cv2.EVENT_MOUSEMOVE:
# pick second point of rect and draw current rect
draw_zoom.rect_in_big_image.br = draw_zoom.shrink_point(x, y)
draw_zoom.reset_big_image()
draw_zoom.update_drawing_big()
elif event == cv2.EVENT_LBUTTONUP:
# finish drawing current rect
draw_zoom.is_drawing_big = False
draw_zoom.rect_in_big_image.br = draw_zoom.shrink_point(x, y)
draw_zoom.rect_in_big_image.regularize()
if draw_zoom.rect_in_big_image.get_area() == 0:
draw_zoom.reset_big_image()
draw_zoom.rect_in_big_image = Rect()
draw_zoom.exist_rect = False
draw_zoom.update_drawing_big()
elif (not draw_zoom.is_drawing_big) and event == cv2.EVENT_RBUTTONDOWN:
# right button down to erase current rect
draw_zoom.rect_in_big_image = Rect()
draw_zoom.exist_rect = False
draw_zoom.reset_big_image()
draw_zoom.update_drawing_big()
else:
# default case: mouse move without rect
draw_zoom.current_pt = (x, y)
draw_zoom.update_drawing_big()
def onmouse_zoom_image(event, x, y, flags, draw_zoom):
if event == cv2.EVENT_LBUTTONDOWN:
# pick first point of rect
draw_zoom.is_drawing_zoom = True
draw_zoom.rect_in_zoom_image.tl = (x, y)
draw_zoom.rect_in_big_image.tl = (x + draw_zoom.zoom_offset[0],
y + draw_zoom.zoom_offset[1])
draw_zoom.exist_rect = True
elif draw_zoom.is_drawing_zoom and event == cv2.EVENT_MOUSEMOVE:
# pick second point of rect and draw current rect
draw_zoom.rect_in_zoom_image.br = (x, y)
draw_zoom.rect_in_big_image.br = draw_zoom.shrink_point(
x + draw_zoom.zoom_offset[0], y + draw_zoom.zoom_offset[1])
draw_zoom.reset_zoom_image()
draw_zoom.reset_big_image()
draw_zoom.update_drawing_zoom()
elif event == cv2.EVENT_LBUTTONUP:
# finish drawing current rect
draw_zoom.is_drawing_zoom = False
draw_zoom.rect_in_big_image.br = draw_zoom.shrink_point(
x + draw_zoom.zoom_offset[0], y + draw_zoom.zoom_offset[1])
draw_zoom.rect_in_big_image.regularize()
if draw_zoom.rect_in_big_image.get_area() == 0:
draw_zoom.reset_big_image()
draw_zoom.rect_in_big_image = Rect()
draw_zoom.rect_in_zoom_image = Rect()
draw_zoom.exist_rect = False
draw_zoom.current_pt = draw_zoom.shrink_point(
x + draw_zoom.zoom_offset[0], y + draw_zoom.zoom_offset[1])
draw_zoom.update_drawing_big()
elif (not draw_zoom.is_drawing_big) and event == cv2.EVENT_RBUTTONDOWN:
# right button down to erase current rect
draw_zoom.rect_in_big_image = Rect()
draw_zoom.rect_in_zoom_image = Rect()
draw_zoom.exist_rect = False
draw_zoom.reset_big_image()
draw_zoom.current_pt = draw_zoom.shrink_point(
x + draw_zoom.zoom_offset[0], y + draw_zoom.zoom_offset[1])
draw_zoom.update_drawing_big()
else:
# mousemove in zoom image will not change the content of image
pass
if __name__ == '__main__':
WIN_NAME_BIG = 'big_image'
WIN_NAME_ZOOM = 'zoom_image'
image = cv2.imread('1.jpg')
draw_zoom = DrawZoom(image, (0, 255, 0))
cv2.namedWindow(WIN_NAME_BIG, 0)
cv2.namedWindow(WIN_NAME_ZOOM, 0)
cv2.setMouseCallback(WIN_NAME_BIG, onmouse_big_image, draw_zoom)
cv2.setMouseCallback(WIN_NAME_ZOOM, onmouse_zoom_image, draw_zoom)
while True:
cv2.imshow(WIN_NAME_BIG, draw_zoom.big_image)
cv2.imshow(WIN_NAME_ZOOM, draw_zoom.zoom_image)
key = cv2.waitKey(30)
if key == 27: # ESC
break
cv2.destroyAllWindows()
结果：
