OpenCV使用CUDA处理图像的教程与实战

重磅干货，第一时间送达

import cv2 as cv

gpu_frame = cv.cuda_GpuMat()

screenshot = cv.imread('media/drip.png')

gpu_frame.upload(screenshot)

gpu_frame.download()

• 在单张图像上使用

• 在多张图像上使用

• 对多张图像使用Dask进行并行延时处理

第1步：上传

import cv2 as cv

gpu_frame = cv.cuda_GpuMat()

screenshot = cv.imread('media/drop.png')

gpu_frame.upload(screenshot)

第2步：处理图像

screenshot = cv.cuda.cvtColor(gpu_frame, cv.COLOR_RGB2BGR)

screenshot = cv.cuda.resize(screenshot, (400, 400))

第3步：下载

screenshot.download()

import cv2 as cv

img_files = ['bear.png', 'drip.png', 'tldr.png', 'frog.png']

# 创建帧来保存图片(cv2.cuda_GpuMat)
gpu_frame = cv.cuda_GpuMat()

for i in range(len(img_files)):
    # 加载图像(CPU)
    screenshot = cv.imread(f"media/{img_files[i]}")

    # 上传到GPU
    gpu_frame.upload(screenshot)
    
    # 转换颜色到opencv (numpy) ndarray→cv2.cuda_GpuMat
    screenshot = cv.cuda.cvtColor(gpu_frame, cv.COLOR_RGB2BGR)
    
    # 反向阈值@ 100
    screenshot = cv.cuda.threshold(screenshot, 105, 255, cv.THRESH_BINARY_INV)
        
    # 调整图像
    screenshot = cv.cuda.resize(screenshot[1], (200, 200))

    # 从GPU下载图像(cv2) cuda_GpuMat→numpy.ndarray
    screenshot = screenshot.download()

import cv2 as cv
import dask.delayed


@dask.delayed
def preprocess(files):
    # 复制图像文件
    i_files = files.copy()
    
    # 创建GPU帧来保存图像
    gpu_frame = cv.cuda_GpuMat()
    
    for i in range(len(i_files)):
        # 加载图像(CPU)
        screenshot = cv.imread(f'media/{i_files[i]}')

        # 上传到GPU
        gpu_frame.upload(screenshot)

        # 转换颜色到opencv (numpy) ndarray→cv2.cuda_GpuMat
        screenshot = cv.cuda.cvtColor(gpu_frame, cv.COLOR_RGB2BGR)
        screenshot = cv.cuda.cvtColor(screenshot, cv.COLOR_BGR2GRAY)

        # 反向阈值@ 100
        screenshot = cv.cuda.threshold(screenshot, 125, 255, cv.THRESH_BINARY)

        # 调整图像
        screenshot = cv.cuda.resize(screenshot[1], (200, 200))
        
        # 从GPU下载图像 (cv2.cuda_GpuMat -> numpy.ndarray)
        screenshot = screenshot.download()

        # 用新图像
        i_files[i] = screenshot
    
    # 输出预处理图像
    return i_files

from dask import compute

img_files = ['bear.png', 'drip.png', 'tldr.png', 'frog.png']
img_files_2 = ['apple.png', 'eye.png', 'window.png', 'blinds.png']

# 设置延迟
set_a = dask.delayed(preprocess)(img_files)
set_b = dask.delayed(preprocess)(img_files_2)

# 开始计算
out_a, out_b = compute(*[set_a, set_b])

进阶阅读：

• https://medium.com/dropout-analytics/opencv-cuda-for-videos-f3dcf346e398
• https://medium.com/dropout-analytics/pycuda-on-jetson-nano-7990decab299
• https://medium.com/dropout-analytics/beginners-guide-to-knn-with-cuml-ddca099f9e9d

参考引用

1. Koriukina, Valeriia. “Getting Started with OpenCV CUDA Module.” Learn OpenCV, Learnopencv.com, 15 Sept. 2020, learnopencv.com/getting-started-opencv-cuda-modul.
2. McWhorter, Paul. “AI on the Jetson Nano LESSON 10: Installing OpenCV for Python 3.” Paul McWhorter — YouTube, Youtube.com/User/Mcwhorpj, 2 Nov. 2019, youtu.be/3QYayL5y2hk.
3. Pulli, Kari; Baksheev, Anatoly; Kornyakov, Kirill; Eruhimov, Victor. “Realtime Computer Vision with OpenCV.” Realtime Computer Vision with OpenCV — ACM Queue, Association for Computing Machinery, 22 Apr. 2012, queue.acm.org/detail.cfm?id=2206309.