Automating censor-bar and mosaic detection with Mask R-CNN

What it does

hent-AI is a computer-vision tool that automates the detection of censor bars and mosaic blurs in illustrated images and video. It trains a Mask R-CNN instance-segmentation model to draw pixel masks around censored regions, feeding those masks into decensoring pipelines like DeepCreamPy or an integrated ESRGAN path for mosaic removal. The result is a GUI-driven detector that replaces the hand-labeling step with inference.

The interesting bit

The author repurposes autonomous-vehicle-grade segmentation—specifically Matterport’s Mask R-CNN—to solve a very specific, very manual content-editing problem. It also includes a synthetic data generator that takes uncensored images, slaps on random bars and mosaics via NudeNet, and spits out polygon annotations, bootstrapping the training set without requiring humans to trace smut by hand.

Key highlights

• Built on Mask R-CNN for instance segmentation of censor regions.
• Supports both bar and mosaic detection, with ESRGAN integration for video decensoring.
• Provides a Windows executable and a Google Colab notebook for cloud-based ESRGAN video processing.
• Includes a synthetic dataset generator (test_combined_generator.py) that auto-annotates training data.
• Requires legacy dependencies (TensorFlow 1.8, Keras 2.2) and CUDA 9.0 for local GPU inference.

Caveats

• The model struggles with white, small, partially-transparent, or fused censor bars, and cannot handle full censorship where the underlying content was never drawn.
• Bar and mosaic modes must be processed separately; mixing both on the same image confuses the pipeline.
• The compiled Windows executable does not support ESRGAN, and RTX cards are known to misbehave with the local ESRGAN path, pushing most video work to the Colab notebook.

Verdict

Worth a look if you are building image-restoration pipelines or curious about niche applications of instance segmentation. Everyone else can safely keep scrolling.