python - 写入同一视频后，视频的 NumPy 数组从原始数组发生变化

Question

我有一个视频(test.mkv)，我已将其转换为 4D NumPy 数组 - (帧、高度、宽度、颜色 channel )。我什至设法将该数组转换回相同的视频( test_2.mkv )而无需更改任何内容。然而，在阅读了这篇新文章后，test_2.mkv ，回到一个新的 NumPy 数组，第一个视频的数组与第二个视频的数组不同，即它们的哈希不匹配，numpy.array_equal()函数返回假。我尝试过同时使用 python-ffmpeg和 scikit-video但无法让数组匹配。
Python-ffmpeg 尝试:

import ffmpeg
import numpy as np
import hashlib

file_name = 'test.mkv'

# Get video dimensions and framerate
probe = ffmpeg.probe(file_name)
video_stream = next((stream for stream in probe['streams'] if stream['codec_type'] == 'video'), None)
width = int(video_stream['width'])
height = int(video_stream['height'])
frame_rate = video_stream['avg_frame_rate']

# Read video into buffer
out, error = (
    ffmpeg
        .input(file_name, threads=120)
        .output("pipe:", format='rawvideo', pix_fmt='rgb24')
        .run(capture_stdout=True)
)

# Convert video buffer to array
video = (
    np
        .frombuffer(out, np.uint8)
        .reshape([-1, height, width, 3])
)

# Convert array to buffer
video_buffer = (
    np.ndarray
        .flatten(video)
        .tobytes()
)

# Write buffer back into a video
process = (
    ffmpeg
        .input('pipe:', format='rawvideo', s='{}x{}'.format(width, height))
        .output("test_2.mkv", r=frame_rate)
        .overwrite_output()
        .run_async(pipe_stdin=True)
)
process.communicate(input=video_buffer)

# Read the newly written video
out_2, error = (
    ffmpeg
        .input("test_2.mkv", threads=40)
        .output("pipe:", format='rawvideo', pix_fmt='rgb24')
        .run(capture_stdout=True)
)

# Convert new video into array
video_2 = (
    np
        .frombuffer(out_2, np.uint8)
        .reshape([-1, height, width, 3])
)

# Video dimesions change
print(f'{video.shape} vs {video_2.shape}') # (844, 1080, 608, 3) vs (2025, 1080, 608, 3)
print(f'{np.array_equal(video, video_2)}') # False

# Hashes don't match
print(hashlib.sha256(bytes(video_2)).digest()) # b'\x88\x00\xc8\x0ed\x84!\x01\x9e\x08 \xd0U\x9a(\x02\x0b-\xeeA\xecU\xf7\xad0xa\x9e\\\xbck\xc3'
print(hashlib.sha256(bytes(video)).digest()) # b'\x9d\xc1\x07xh\x1b\x04I\xed\x906\xe57\xba\xf3\xf1k\x08\xfa\xf1\xfaM\x9a\xcf\xa9\t8\xf0\xc9\t\xa9\xb7'

Scikit 视频尝试:

import skvideo.io as sk
import numpy as np

video_data = sk.vread('test.mkv')

sk.vwrite('test_2_ski.mkv', video_data)

video_data_2 = sk.vread('test_2_ski.mkv')

# Dimensions match but...
print(video_data.shape) # (844, 1080, 608, 3)
print(video_data_2.shape) # (844, 1080, 608, 3)

# ...array elements don't
print(np.array_equal(video_data, video_data_2)) # False

# Hashes don't match either
print(hashlib.sha256(bytes(video_2)).digest()) # b'\x8b?]\x8epD:\xd9B\x14\xc7\xba\xect\x15G\xfaRP\xde\xad&EC\x15\xc3\x07\n{a[\x80'
print(hashlib.sha256(bytes(video)).digest()) # b'\x9d\xc1\x07xh\x1b\x04I\xed\x906\xe57\xba\xf3\xf1k\x08\xfa\xf1\xfaM\x9a\xcf\xa9\t8\xf0\xc9\t\xa9\xb7'

我不明白我哪里出错了，两个各自的文档都没有强调如何完成这个特定的任务。任何帮助表示赞赏。谢谢你。

Answer 1

写入和读取视频文件时需要获得相同的哈希值，需要特别注意。
在比较哈希之前，请先尝试查看视频。
执行您的代码给了我以下输出(video_2 的第一帧):

当输入(视频的第一帧)是:

我建议进行以下修改:

使用 AVI 容器(而不是 MKV)来存储 test_2原始视频格式的视频。
AVI 视频容器最初是为存储原始视频而设计的。
可能有一种方法可以在 MKV 容器中存储原始或无损 RGB 视频，但我不知道这种选项。

设置test_2的输入像素格式视频。
添加参数:pixel_format='rgb24' .
注意:我修改为pixel_format='bgr24' , 因为 AVI 支持 bgr24而不是 rgb24 .

为 test_2 选择视频无损编解码器视频。
您可以选择vcodec='rawvideo' (AVI 支持 rawvideo 编解码器，但 MKV 不支持)。

笔记:
要获得相等的哈希，您需要寻找支持 rgb24 的无损视频编解码器(或 bgr24 )像素格式。
大多数无损编解码器将像素格式从 RGB 转换为 YUV。
RGB 到 YUV 的转换有舍入错误，防止相等的散列。
(我想有办法绕过它，但它有点复杂)。

这是您的完整代码，只做了一些修改:

import ffmpeg
import numpy as np
import hashlib

file_name = 'test.mkv'

# Get video dimensions and framerate
probe = ffmpeg.probe(file_name)
video_stream = next((stream for stream in probe['streams'] if stream['codec_type'] == 'video'), None)
width = int(video_stream['width'])
height = int(video_stream['height'])
frame_rate = video_stream['avg_frame_rate']

# Read video into buffer
out, error = (
    ffmpeg
        .input(file_name, threads=120)
        .output("pipe:", format='rawvideo', pix_fmt='bgr24')  # Select bgr24 instead of rgb24 (becasue raw AVI requires bgr24).
        .run(capture_stdout=True)
)

# Convert video buffer to array
video = (
    np
        .frombuffer(out, np.uint8)
        .reshape([-1, height, width, 3])
)

# Convert array to buffer
video_buffer = (
    np.ndarray
        .flatten(video)
        .tobytes()
)

# Write buffer back into a video
process = (
    ffmpeg
        .input('pipe:', format='rawvideo', s='{}x{}'.format(width, height), pixel_format='bgr24', r=frame_rate)  # Set input pixel format.
        .output("test_2.avi", vcodec='rawvideo')  # Select video code "rawvideo"
        .overwrite_output()
        .run_async(pipe_stdin=True)
)
process.communicate(input=video_buffer)

# Read the newly written video
out_2, error = (
    ffmpeg
        .input("test_2.avi", threads=40)
        .output("pipe:", format='rawvideo', pix_fmt='bgr24')
        .run(capture_stdout=True)
)

# Convert new video into array
video_2 = (
    np
        .frombuffer(out_2, np.uint8)
        .reshape([-1, height, width, 3])
)

# Video dimesions change
print(f'{video.shape} vs {video_2.shape}') # (844, 1080, 608, 3) vs (844, 1080, 608, 3)
print(f'{np.array_equal(video, video_2)}') # True

# Hashes do match
print(hashlib.sha256(bytes(video_2)).digest())
print(hashlib.sha256(bytes(video)).digest())

结果(相同的哈希): True b"\xd1yy\x97\x8e\xce\x13\xbcI#\xd2PMP\x80(i+5\xe1\xcd\xab\xf3f\xbe\xcd\xd5'\xbaq\xdd\x9b" b"\xd1yy\x97\x8e\xce\x13\xbcI#\xd2PMP\x80(i+5\xe1\xcd\xab\xf3f\xbe\xcd\xd5'\xbaq\xdd\x9b"
更新:
使用 ffv1 编码器:
使用 ffv1 实现相同的哈希值.mkv 编码器

选择 vcodec='ffv1'在 output() 的论点中.

还有一件事:

移动参数 r=frame_rate从输出参数到 输入 论据。
这不直观......但是当从帧创建视频时，帧速率应该定义为输入的参数。

 # Write buffer back into a video
 process = (
     ffmpeg
         .input('pipe:', format='rawvideo', s='{}x{}'.format(width, height), pixel_format='rgb24', r=frame_rate)  # Set input pixel format.
         .output("test_2.mkv", vcodec='ffv1')  # Select video code "rawvideo"
         .overwrite_output()
         .run_async(pipe_stdin=True)
 )

这是完整的代码示例:

import ffmpeg
import numpy as np
import hashlib

file_name = 'test.mkv'

# Get video dimensions and framerate
probe = ffmpeg.probe(file_name)
video_stream = next((stream for stream in probe['streams'] if stream['codec_type'] == 'video'), None)
width = int(video_stream['width'])
height = int(video_stream['height'])
frame_rate = video_stream['avg_frame_rate']

# Read video into buffer
out, error = (
    ffmpeg
        .input(file_name, threads=120)
        .output("pipe:", format='rawvideo', pix_fmt='rgb24')  # Select rgb24 instead of rgb24 (becasue raw AVI requires rgb24).
        .run(capture_stdout=True)
)

# Convert video buffer to array
video = (
    np
        .frombuffer(out, np.uint8)
        .reshape([-1, height, width, 3])
)

# Convert array to buffer
video_buffer = (
    np.ndarray
        .flatten(video)
        .tobytes()
)

# Write buffer back into a video
process = (
    ffmpeg
        .input('pipe:', format='rawvideo', s='{}x{}'.format(width, height), pixel_format='rgb24', r=frame_rate)  # Set input pixel format.
        .output("test_2.mkv", vcodec='ffv1')  # Select video code "rawvideo"
        .overwrite_output()
        .run_async(pipe_stdin=True)
)
process.communicate(input=video_buffer)

# Read the newly written video
out_2, error = (
    ffmpeg
        .input("test_2.mkv", threads=40)
        .output("pipe:", format='rawvideo', pix_fmt='rgb24')
        .run(capture_stdout=True)
)

# Convert new video into array
video_2 = (
    np
        .frombuffer(out_2, np.uint8)
        .reshape([-1, height, width, 3])
)

# Video dimesions change
print(f'{video.shape} vs {video_2.shape}') # (844, 1080, 608, 3) vs (844, 1080, 608, 3)
print(f'{np.array_equal(video, video_2)}') # True

# Hashes do match
print(hashlib.sha256(bytes(video_2)).digest())
print(hashlib.sha256(bytes(video)).digest())

结果(相同的哈希，使用您的输入文件): True b'\x9d\xc1\x07xh\x1b\x04I\xed\x906\xe57\xba\xf3\xf1k\x08\xfa\xf1\xfaM\x9a\xcf\xa9\t8\xf0\xc9\t\xa9\xb7' b'\x9d\xc1\x07xh\x1b\x04I\xed\x906\xe57\xba\xf3\xf1k\x08\xfa\xf1\xfaM\x9a\xcf\xa9\t8\xf0\xc9\t\xa9\xb7'
更新:
使用 Scikit 视频 :
以下代码示例使用 Scikit-Video。
我找不到选择 ffv1 的方法编解码器，不使用 skvideo.io.FFmpegWriter .
该实现使用 for 循环逐帧写入视频。

import skvideo.io as sk
import numpy as np
import hashlib

video_data = sk.vread('test.mkv')

# Create FFmpeg vidoe writer
writer = sk.FFmpegWriter('test_2_ski.mkv', outputdict={'-vcodec': 'ffv1' })

#sk.vwrite('test_2_ski.mkv', video_data)

# Write frame by frame in a loop
for i in range(video_data.shape[0]):
    writer.writeFrame(video_data[i, :, :, :])

writer.close()  # Close video writer.

video_data_2 = sk.vread('test_2_ski.mkv')

# Dimensions match
print(video_data.shape) # (844, 1080, 608, 3)
print(video_data_2.shape) # (844, 1080, 608, 3)

# Array elements match
print(np.array_equal(video_data, video_data_2))

# Hashes match
print(hashlib.sha256(bytes(video_data_2)).digest())
print(hashlib.sha256(bytes(video_data)).digest())