使用 matplotlib 和 for 循环保存许多图像的快速方法

Question

我注意到以下两种使用 matplotlib 加载和保存图像的解决方案之间存在很大的性能差距。谁能解释为什么，以及在 python for 循环中保存图像的最佳（和最快）方法是什么？

实施 1： 在 for 循环之外创建一个图形，更新显示的内容然后保存。

fig, a = plt.subplots(1, 3, figsize=(30, 20))  # <--------

# list_of_fnames is just a list of file names
for k, fname in enumerate(list_of_fnames):
    with Image.open(fname) as img:
        x = np.array(img)

    y = process_image_fn1(x)
    z = process_image_fn2(x)

    a[0].imshow(x)
    a[1].imshow(y)
    a[2].imshow(z)

    output_filename = f'results_{k}.png'
    plt.savefig(output_filename, dpi=320, format='png', transparent=False, bbox_inches='tight', pad_inches=0)

实施 2： 在 for 循环中创建一个图形，保存它，最后销毁它。

# list_of_fnames is just a list of file names
for k, fname in enumerate(list_of_fnames):
    with Image.open(fname) as img:
        x = np.array(img)

    y = process_image_fn1(x)
    z = process_image_fn2(x)

    fig, a = plt.subplots(1, 3, figsize=(30, 20))  # <--------
    a[0].imshow(x)
    a[1].imshow(y)
    a[2].imshow(z)

    output_filename = f'results_{k}.png'
    plt.savefig(output_filename, dpi=320, format='png', transparent=False, bbox_inches='tight', pad_inches=0)

    plt.close()  # <--------

Answer 1

第一个选项可以通过多种方式进行改进。

1. 删除之前绘制的 AxesImages（来自 imshow），这样您就不会不断增加轴上绘制的图像数量

fig, a = plt.subplots(1, 3, figsize=(30, 20))  # <--------

# list_of_fnames is just a list of file names
for k, fname in enumerate(list_of_fnames):

    for ax in a:
        ax.images.pop()

    with Image.open(fname) as img:
        x = np.array(img)

    y = process_image_fn1(x)
    z = process_image_fn2(x)

    a[0].imshow(x)
    a[1].imshow(y)
    a[2].imshow(z)

    output_filename = f'results_{k}.png'
    plt.savefig(output_filename, dpi=320, format='png', transparent=False, bbox_inches='tight', pad_inches=0)

1. 或者，为每个轴创建一次 AxesImages，而不是每次迭代都重新绘制它们，而是使用 .set_array() 更改 AxesImage 上绘制的内容

fig, a = plt.subplots(1, 3, figsize=(30, 20))  # <--------

# list_of_fnames is just a list of file names
for k, fname in enumerate(list_of_fnames):
    with Image.open(fname) as img:
        x = np.array(img)

    y = process_image_fn1(x)
    z = process_image_fn2(x)

    if k == 0:
        im0 = a[0].imshow(x)
        im1 = a[1].imshow(y)
        im2 = a[2].imshow(z)
    else:
        im0.set_array(x)
        im1.set_array(y)
        im2.set_array(z)

    output_filename = f'results_{k}.png'
    plt.savefig(output_filename, dpi=320, format='png', transparent=False, bbox_inches='tight', pad_inches=0)