如何用GPU加速OpenCV Python的imread/imwrite及性能优化疑问

核心背景与GPU加速疑问 #

在Ubuntu系统上使用OpenCV Python 4.8.1进行图片分割操作，当前性能表现不佳，希望优化提速。已下载OpenCV 4.8.1源码并完成编译，通过cv2.cuda.getCudaEnabledDeviceCount()确认GPU计数为1，但不清楚如何让imread/imwrite调用GPU进行加速。

我的OpenCV Python代码如下：

# Load and split the image using OpenCV
image = cv2.imread(request_data.file_location)
height, width, _ = image.shape
overlap = 100  # Adjust the overlap size as needed
size = 500

results = []
for i in range(0, height, size - overlap):
    for j in range(0, width, size - overlap):
        img_part = image[i:i + size, j:j + size]
        output_file = os.path.join(output_dir, f"result_{i}_{j}.jpg")
        cv2.imwrite(output_file, img_part)
        results.append(output_file)

return {"path": results}

多进程优化可行性疑问 #

若无法通过GPU加速imread/imwrite（当前仅占用1核CPU），是否可以使用多进程优化？另外，多进程是不是无法提升imread()的性能？

Pillow-SIMD性能结果疑问 #

尝试了Pillow-SIMD 9.0.0.post1并将压缩级别设为1，但性能比OpenCV更慢：OpenCV读取耗时2秒、写入耗时2秒；Pillow-SIMD读取耗时1秒、写入耗时3秒。这一结果是否合理？我查阅资料发现多数场景下Pillow-SIMD比OpenCV更快。

我的Pillow代码如下：

image = Image.open(request_data.file_location)

# Define split parameters
size = 500
overlap = 100

results = []

for i in range(0, image.height, size - overlap):
    for j in range(0, image.width, size - overlap):
        # Crop and save the image using Pillow
        img_part = image.crop((j, i, j + size, i + size))
        output_file = os.path.join(output_dir, f"result_{i}_{j}.jpg")
        img_part.save(output_file)

        results.append(output_file)

内容的提问来源于stack exchange，提问作者Dennys