GitLocker: The Coding Marketplace

Description:

cvlayer 0.32.0

cvlayer

OpenCV Layer Helper
Overview
When performing image-based machine learning/inference, many situations arise that require pre-/post-processing appropriate for the domain environment.
A commonly used library today is OpenCV.
Image processing with OpenCV requires insight into the image and the domain.
However, gaining these insights often requires doing something far from the essence. (Visualization, parameter modification, library conflicts, assertions, etc...)
Here are some problems that occurred to me:
...
Features
...
Installation
Install cvlayer:
pip install cvlayer

Install cvlayer with opencv-python:
pip install cvlayer[opencv]

Install cvlayer with opencv-python-headless:
pip install cvlayer[headless]

Usage
CvLayer
Just inherit cvlayer.CvLayer.
from cvlayer import CvLayer

class YourApp(CvLayer):
def func(self, image):
self.cvl_cvt_color_bgr2hsv(image)

CvWindow
Just inherit cvlayer.CvWindow.
from cvlayer import CvWindow
from numpy.typing import NDArray

class YourApp(CvWindow):
def on_frame(self, image: NDArray) -> NDArray:
# TODO: Implement image processing logic ...
return image

YourApp("/path/to/video/file.mp4").run()

Example
The following sample is a Perspective Transform example:
from sys import argv, stderr
from sys import exit as sys_exit
from typing import List, Optional

from cvlayer import CvLayer, CvMixin, CvWindow
from cvlayer.typing import PointI
from cvlayer.palette.basic import RED
from numpy.typing import NDArray

class YourApp(CvWindow, CvMixin, CvLayer):
_points: List[PointI]

def __init__(self, source: str, destination: Optional[str] = None):
super().__init__(source, destination)
left_top = 930, 2750
left_bottom = 846, 3098
right_top = 1091, 2750
right_bottom = 1361, 3098
self._points = [left_top, left_bottom, right_top, right_bottom]
self._scale = 2, 4

def on_frame(self, image: NDArray) -> NDArray:
with self.layer("select-roi") as layer:
self.roi = layer.param("roi").build_select_roi().value
layer.frame = layer.prev_frame

with self.layer("perspective-points") as layer:
points = layer.param("pp").build_select_points(self._points).value
canvas = layer.prev_frame.copy()
for p in points:
self.cvl_draw_point(canvas, p, color=RED)
layer.frame = canvas

with self.layer("perspective-transform") as layer:
sw = layer.param("scale-width").build_uint(self._scale[0]).value
sh = layer.param("scale-height").build_uint(self._scale[1]).value

xs = list(map(lambda point: point[0], self._points))
ys = list(map(lambda point: point[1], self._points))
x1, x2 = min(xs), max(xs)
y1, y2 = min(ys), max(ys)
width, height = abs(x2 - x1) * sw, abs(y2 - y1) * sh
roi = 0, 0, width, height
m = self.cvl_get_perspective_transform_with_quadrilateral(
left_top=points[0],
left_bottom=points[1],
right_top=points[2],
right_bottom=points[3],
destination_roi=roi,
)
layer.frame = self.cvl_warp_perspective(image, m, (width, height))

with self.layer("hsv") as layer:
layer.frame = hsv = self.cvl_cvt_color_bgr2hsv(layer.prev_frame)
with self.layer("hsv-h") as layer:
layer.frame = h = hsv[:, :, 0]
with self.layer("hsv-s") as layer:
layer.frame = s = hsv[:, :, 1]
with self.layer("hsv-v") as layer:
layer.frame = v = hsv[:, :, 2]

assert h is not None
assert s is not None
assert v is not None

self.cvm_gaussian_blur("v-blur", (3, 19), 0.0, 7.0)
self.cvm_threshold_binary("v-thresh", 230)

return self.last_frame

def test_main(*args) -> None:
source = args[1]
destination = args[2] if len(args) >= 3 else None

try:
app = YourApp(source, destination)
app.run()
except Exception as e:
print(e, file=stderr)
sys_exit(1)
else:
sys_exit(0)

if __name__ == "__main__":
test_main(*argv)

License
See the LICENSE file for details. In summary,
cvlayer is licensed under the MIT license.