a312863063/Video-Auto-Wipe

Video-Auto-Wipe is a deep learning project for automatically removing unwanted, fixed-pattern content from videos—specifically subtitles, static logos, and dynamic/animated icons. It is a genuine computer vision / video inpainting project.

What it does: The model automatically detects and erases specific types of visual noise from video frames. It identifies subtitles by their uniform text style, static logos by pixels that remain motionless over time, and dynamic icons by flickering or moving fixed pixel blocks. The project currently open-sources the subtitle removal model, while the logo and dynamic icon removal models are demonstrated but not yet released.

How it works: The training process uses a two-step approach. First, the model is trained to perceive the target foreground (e.g., subtitles) in the time domain. Second, this perception module is fused into an erasing (inpainting) model for end-to-end fine-tuning. To generate training data, the authors created synthetic foregrounds—using Python's ImageDraw library to generate random text and pixel blocks, and Adobe Premiere Pro for dynamic effects—and overlaid them on real video backgrounds (2,709 movie clips and 864 variety show clips).

Key features:

• Open-sourced subtitle removal model with pretrained weights and a simple demo.py script.
• Demonstrates effective erasure of static and dynamic logos (though those specific weights are not shared).
• Provides large synthetic training datasets for background and foreground generation.

Tech stack:

• Python, PyTorch (requires torch>1.0)
• ImageDraw for synthetic data generation

Target audience: Developers and researchers interested in video editing, computer vision, video inpainting, or AIGC (AI-Generated Content) tools.

Notable limitations:

• Only the subtitle removal model is open-sourced; logo and dynamic icon models are not.
• Pretrained weights and datasets are hosted on Baidu Pan (Chinese cloud storage).
• The README is minimal regarding exact dependency installation (suggests installing missing packages as needed).