AgentSkillsCN

Video Processing

Video Processing

中文原作
SKILL.md

视频处理最佳实践

概述

视频处理开发最佳实践，包含 FFmpeg、OpenCV 和关键帧提取。

适用场景

  • •视频分析和处理
  • •关键帧提取
  • •图像预筛选
  • •视频转码

核心技术栈

  • •ffmpeg-python: FFmpeg 封装
  • •opencv-python: 图像处理
  • •Pillow: 图像操作

FFmpeg 基础

提取音频

python
import ffmpeg

ffmpeg.input("video.mp4").output("audio.wav", vn=None, acodec="pcm_s16le", ar=16000, ac=1).run()

提取关键帧

python
ffmpeg.input("video.mp4").filter('fps', fps=1/30).output("frame_%04d.jpg").run()

获取视频信息

python
probe = ffmpeg.probe("video.mp4")
video_info = next(s for s in probe['streams'] if s['codec_type'] == 'video')
duration = float(video_info['duration'])
width = int(video_info['width'])
height = int(video_info['height'])

OpenCV 基础

读取和显示

python
import cv2

# 读取视频
cap = cv2.VideoCapture("video.mp4")

while True:
    ret, frame = cap.read()
    if not ret:
        break
    # 处理帧
    
cap.release()

图像预筛选

python
def is_interesting_frame(frame):
    """判断帧是否包含有价值内容"""
    gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    
    # 边缘检测
    edges = cv2.Canny(gray, 50, 150)
    edge_ratio = np.count_nonzero(edges) / edges.size
    
    # 变化率检测（对比相邻帧）
    # ...
    
    return edge_ratio > threshold

直方图比较

python
def compare_frames(frame1, frame2):
    """比较两帧的相似度"""
    hist1 = cv2.calcHist([frame1], [0], None, [256], [0, 256])
    hist2 = cv2.calcHist([frame2], [0], None, [256], [0, 256])
    return cv2.compareHist(hist1, hist2, cv2.HISTCMP_CORREL)

关键帧提取策略

基于时间间隔

python
def extract_keyframes(video_path, interval_sec=30):
    """按时间间隔提取关键帧"""
    cap = cv2.VideoCapture(video_path)
    fps = cap.get(cv2.CAP_PROP_FPS)
    interval_frames = int(fps * interval_sec)
    
    frames = []
    frame_count = 0
    
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        if frame_count % interval_frames == 0:
            frames.append(frame)
        frame_count += 1
    
    cap.release()
    return frames

基于场景变化

python
def extract_scenes(video_path, threshold=0.7):
    """基于场景变化提取关键帧"""
    cap = cv2.VideoCapture(video_path)
    prev_frame = None
    keyframes = []
    
    while True:
        ret, frame = cap.read()
        if not ret:
            break
            
        if prev_frame is not None:
            similarity = compare_frames(prev_frame, frame)
            if similarity < threshold:  # 场景变化
                keyframes.append(frame)
        else:
            keyframes.append(frame)
            
        prev_frame = frame
    
    cap.release()
    return keyframes

待完善

  • • 添加完整处理流程示例
  • • 补充性能优化技巧
  • • 补充多线程处理
  • • 补充 GPU 加速