Extract frames and OCR text from MP4 videos

Video Content Extractor turns on-screen text from an MP4 file into a structured Markdown report. It gathers video metadata, captures frames at a configurable time interval, runs OCR on each image, and organizes the recognized text with timestamps. It is designed for Codex CLI and performs the processing locally.

The metadata stage uses ffprobe to collect duration, resolution, frame rate, codec information, and file size. FFmpeg then creates timestamped JPEG frames. Tesseract reads each frame, with a fallback to fully automatic page segmentation if the default segmentation mode produces no meaningful text. The final report combines metadata and frame-by-frame text rather than presenting an unlocated block of transcription.

Use it for text in video presentations, lectures, and screencasts, especially when no embedded subtitles are available. It is also relevant to requests to read what appears on screen or create searchable text from a video file.

This workflow does not extract audio or perform speech-to-text transcription. Its frame selection is based on elapsed time rather than scene changes, so similar neighboring frames can be produced. OCR quality depends on video resolution, text size, and font clarity. The recognized text therefore reflects what can be read from the sampled images.

FFmpeg and Tesseract must be installed and accessible through PATH. Large videos sampled at short intervals can create many images, making available disk space and processing time practical constraints. Increasing the interval reduces the number of frames processed, with a corresponding change in how densely the video is sampled.

Inputs are the video path, output directory, sampling interval in seconds, and OCR language choice. The default frame interval is 30 seconds. Use shorter intervals of roughly 10 to 15 seconds for frequent text changes and longer intervals of 30 to 60 seconds for slow lectures or slides to reduce repetition.

Outputs include a Markdown report and a directory of timestamped frames. The basic lecture example produces lecture.md and lecture_frames. An English-only custom-interval example is:

python scripts/extract_video.py recording.mp4 ./output 60 eng

The bilingual example uses Chinese and English recognition. Chinese content requires the chi_sim language pack. If OCR returns garbled text, the workflow checks language-pack installation with Tesseract's language-list command. If FFmpeg cannot be found, it checks PATH and the installed version.

The original video is not modified or deleted. The skill only reads the source file and writes extracted images and reports. It sends no data over the network, and the processing tools use fixed, pre-vetted arguments.

The implementation combines FFmpeg for frame extraction, ffprobe for metadata, and Tesseract OCR for text recognition. Language options include eng and chi_sim+eng. Related skills cover broader media summaries using visual and audio cues, and OCR on static images or scanned documents.

It suits people who need searchable on-screen text from recorded presentations, lectures, or screencasts. It is useful when local processing and timestamped references are required, and the goal is visual text rather than spoken content.