Python+OpenCV讀寫視頻的方法詳解

Posted on by

讀視頻，提取幀

接口函數：cv2.VideoCapture()

通過video_capture = cv2.VideoCapture(video_path)可以獲取讀取視頻的句柄。而後再通過flag, frame = video_capture.read()可以讀取當前幀，flag表示讀取是否成功，讀取成功後，句柄會自動移動到下一幀的位置。讀取結束後使用video_capture.release()釋放句柄。

一個簡單的逐幀讀取的程序如下：

import cv2

video_capture = cv2.VideoCapture(video_path)
while True:
    flag, frame = video_capture.read()
    if not flag:
        break
    # do something with frame
video_capture.release()

獲取視頻信息

為瞭能更好更靈活地瞭解並讀取視頻，我們有時候需要獲取視頻的一些信息，比如幀率，總幀數等等。獲取這些信息的方法是調用video_capture.get(PROP_ID)方法，其中PROP_ID是OpenCV定義的一些常量。

常用的信息及示例如下：

import cv2

video_path = r'D:\peppa\Muddy_Puddles.mp4'
video_capture = cv2.VideoCapture(video_path)

frame_num = video_capture.get(cv2.CAP_PROP_FRAME_COUNT) # ==> 總幀數
fps = video_capture.get(cv2.CAP_PROP_FPS)               # ==> 幀率
width = video_capture.get(cv2.CAP_PROP_FRAME_WIDTH)     # ==> 視頻寬度
height = video_capture.get(cv2.CAP_PROP_FRAME_HEIGHT)   # ==> 視頻高度
pos = video_capture.get(cv2.CAP_PROP_POS_FRAMES)        # ==> 句柄位置

video_capture.set(cv2.CAP_PROP_POS_FRAMES, 1000)        # ==> 設置句柄位置
pos = video_capture.get(cv2.CAP_PROP_POS_FRAMES)        # ==> 此時 pos = 1000.0

video_capture.release()

句柄位置指的是下一次調用read()方法讀取到的幀號，幀號索引從0開始。

使用set(cv2.CAP_PROP_POS_FRAMES)讀取指定幀

從上面代碼中可以看到我們使用瞭set方法來設置句柄的位置，這個功能在讀取指定幀時很有用，這樣我們不必非要使用read()遍歷到指定位置。

但問題來瞭，這種方式讀取到的內容和read()遍歷讀取到的內容是否完全相同？

做個簡單的實驗，下面用兩種方法分別讀取同一個視頻的[100, 200)幀，然後檢查讀取的內容是否完全相同，結果是True。

import cv2
import numpy as np

video_path = r'D:\peppa\Muddy_Puddles.mp4'
video_capture = cv2.VideoCapture(video_path)
cnt = -1
frames1 = []
while True:
    cnt += 1
    flag, frame = video_capture.read()
    assert flag
    if 100 <= cnt < 200:
        frames1.append(frame)
    if cnt >= 200:
        break
video_capture.release()

video_capture = cv2.VideoCapture(video_path)
frames2 = []
for i in range(100, 200):
    video_capture.set(cv2.CAP_PROP_POS_FRAMES, i)
    flag, frame = video_capture.read()
    assert flag
    frames2.append(frame)
video_capture.release()

frames1 = np.array(frames1)
frames2 = np.array(frames2)
print(np.all(frames1 == frames2))  # ==> check whether frames1 is same as frames2, result is True

接下來看看利用set讀取的效率。還是利用小豬佩奇第一集做實驗，這個視頻共7788幀，下面分別用兩種方法遍歷讀取視頻中所有幀。第二種方法明顯比第一種慢得多，所以這就很苦逼瞭。。。如果幀間隔比較小的話，單純用read()進行遍歷效率高；如果幀間隔比較大的話，用set()設置位置，然後read()讀取效率高。

（如果給第二種方法加個判斷，每隔n幀讀取一次，那麼效率確實會提高n倍，可以自行嘗試）

import cv2
import numpy as np
import time

video_path = r'D:\peppa\Muddy_Puddles.mp4'
video_capture = cv2.VideoCapture(video_path)
t0 = time.time()
while True:
    flag, frame = video_capture.read()
    if not flag:
        break
t1 = time.time()
video_capture.release()

video_capture = cv2.VideoCapture(video_path)
t2 = time.time()
frame_num = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))
for i in range(frame_num):
    video_capture.set(cv2.CAP_PROP_POS_FRAMES, i)
    flag, frame = video_capture.read()
    assert flag
t3 = time.time()
video_capture.release()

print(t1 - t0)  # ==> 76.3 s
print(t3 - t2)  # ==> 345.1 s

讀取函數（重點）

上面我們使用瞭兩種方法讀取視頻幀，第一種是使用read()進行暴力遍歷，第二種是使用set()設置幀號，再使用read()讀取。兩種方法讀取到的結果完全一樣，但是效率在不同的情況下各有優勢，所以為瞭最大化發揮兩者的優勢，在寫讀取幀函數時，就要把兩種方式都寫進去，由參數來決定使用哪種模式，這樣用戶可以針對電腦的硬件做一些簡單實驗後自行決定。

# -*- coding: utf-8 -*-
import os
import cv2


def _extract_frame_mode_1(video_capture, frame_list, root_folder, ext='png'):
    """
    extract video frames and save them to disk. this method will go through all
    the frames using video_capture.read()

    Parameters:
    -----------
    video_capture: obtained by cv2.VideoCapture()
    frame_list: list
        list of frame numbers
    root_folder: str
        root folder to save frames
    ext: str
        extension of filename
    """
    frame_list = sorted(frame_list)
    video_capture.set(cv2.CAP_PROP_POS_FRAMES, 0)
    cnt = -1
    index = 0
    while True:
        cnt += 1
        flag, frame = video_capture.read()
        if not flag:
            break
        if cnt == frame_list[index]:
            filename = os.path.join(root_folder, str(cnt) + '.' + ext)
            cv2.imwrite(filename, frame)
            index += 1


def _extract_frame_mode_2(video_capture, frame_list, root_folder, ext='png'):
    """
        extract video frames and save them to disk. this method will use
        video_capture.set() to locate the frame position and then use
        video_capture.read() to read

        Parameters:
        -----------
        video_capture: obtained by cv2.VideoCapture()
        frame_list: list
            list of frame numbers
        root_folder: str
            root folder to save frames
        ext: str
            extension of image filename
        """
    for i in frame_list:
        video_capture.set(cv2.CAP_PROP_POS_FRAMES, i)
        flag, frame = video_capture.read()
        assert flag
        filename = os.path.join(root_folder, str(i) + '.' + ext)
        cv2.imwrite(filename, frame)


def extract_frame(video_path, increment=None, frame_list=None,
                  mode=1, ext='png'):
    """
    extract video frames and save them to disk. the root folder to save frames
    is same as video_path (without extension)
    
    Parameters:
    -----------
    video_path: str
        video path
    increment: int of 'fps'
        increment of frame indexes
    frame_list: list
        list of frame numbers
    mode: int, 1 or 2
        1: go through all the frames using video_capture.read()
        2: use video_capture.set() to locate the frame position and then use
        video_capture.read() to read
    ext: str
        extension of image filename
    """
    video_capture = cv2.VideoCapture(video_path)
    frame_num = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))

    if increment is None:
        increment = 1
    elif increment == 'fps':
        fps = video_capture.get(cv2.CAP_PROP_FPS)
        increment = round(fps)

    if frame_list is None:
        frame_list = [i for i in range(0, frame_num, increment)]

    if frame_num // len(frame_list) > 5 and mode == 1:
        print("the frames to be extracted is too sparse, "
              "please consider setting mode = 2 to accelerate")

    root_folder = os.path.splitext(video_path)[0]
    os.makedirs(root_folder, exist_ok=True)
    if mode == 1:
        _extract_frame_mode_1(video_capture, frame_list, root_folder, ext)
    elif mode == 2:
        _extract_frame_mode_2(video_capture, frame_list, root_folder, ext)
    video_capture.release()


if __name__ == '__main__':
    video_path = r'D:\peppa\Muddy_Puddles.mp4'
    extract_frame(video_path, increment=30, mode=2)

將圖像寫為視頻

寫視頻沒有那麼多需要註意的地方，主要使用的接口函數是cv2.VideoWriter(video_path, fourcc, fps, size)，該函數的主要註意點是入參的設置，video_path是輸出視頻的文件名，fps是幀率，size是視頻的寬高，待寫入視頻的圖像的尺寸必需與size一致。其中不太容易理解的是與視頻編碼相關的fourcc，該參數的設置需要使用另外一個接口函數：cv2.VideoWriter_fourcc(c1, c2, c3, c4)，c1-c4分別是四個字符。

示例

因為獲取圖像的方式多種多樣，而寫視頻又比較簡單，所以不太適合將這部分寫成函數，下面以一個例子呈現。

video_path = r'D:\peppa\Muddy_Puddles.avi'
root_folder = r'D:\peppa\Muddy_Puddles'

fourcc = cv2.VideoWriter_fourcc('X', 'V', 'I', 'D')
fps = 25
size = (1920, 1080)

video_writer = cv2.VideoWriter(video_path, fourcc, fps, size)
for i in range(0, 7788, 30):
    filename = os.path.join(root_folder, str(i) + '.png')
    image = cv2.imread(filename)
    video_writer.write(image)
video_writer.release()

fourcc

fourcc有時候需要多嘗試一下，因為不同電腦裡安裝的編解碼器可能不太一樣，不見得隨便設置一個參數就一定能成功，fourcc有非常多，比如：

paramters codec extension
(‘P’,‘I’,‘M’,‘1’) MPEG-1 avi
(‘M’,‘J’,‘P’,‘G’) motion-jpeg mp4
(‘M’,‘P’,‘4’,‘V’) MPEG-4 mp4
(‘X’,‘2’,‘6’,‘4’) H.264 mp4
(‘M’, ‘P’, ‘4’, ‘2’) MPEG-4.2  
(‘D’, ‘I’, ‘V’, ‘3’)  MPEG-4.3  
(‘D’, ‘I’, ‘V’, ‘X’) MPEG-4 avi
(‘U’, ‘2’, ‘6’, ‘3’) H263  
(‘I’, ‘2’, ‘6’, ‘3’)  H263I flv
(‘F’, ‘L’, ‘V’, ‘1’)  FLV1  
(‘X’,‘V’,‘I’,‘D’)  MPEG-4 avi
(‘I’,‘4’,‘2’,‘0’)  YUV avi

上表中的後綴名似乎並不需要嚴格遵守。

以上就是Python+OpenCV讀寫視頻的方法詳解的詳細內容，更多關於Python OpenCV讀寫視頻的資料請關註WalkonNet其它相關文章！

推薦閱讀：