CNN-only EfficientUnMarker

A fast, lightweight, and practical watermark removal pipeline using only a CNN-based detector (MobileNetV3).
This repository provides a minimal, reproducible baseline for universal watermark removal with a focus on speed, memory efficiency, and simplicity.

Overview

• Purpose: Remove visible watermarks using a 3-stage pipeline (frequency analysis → multi-scale attack → quality safeguard) guided only by a CNN-based detector.
• Target hardware: RTX 5070 Ti–class GPU or similar.
• Design goals: <10s per image, <8GB VRAM, >90% removal success.
• Relation to EfficientUnMarker: Distilled version of the full EfficientUnMarker project, containing only the CNN-based (MobileNetV3) detection and attack components.

Features

• Stage 1 – Frequency analysis: torch.fft-based spectrum analysis + MobileNetV3 features to generate a frequency band mask.
• Stage 2 – Multi-scale attack: 256→original resolution optimization with early stopping and a soft PSNR penalty to avoid excessive degradation.
• Stage 3 – Optional refinement: Additional quality refinement stage (disabled by default in this baseline).
• Metrics: Success flag, detection score, PSNR, runtime, and peak VRAM usage.
• Configurable: YAML-based configuration for attack hyperparameters and stopping criteria.

Dataset

Evaluation uses the Large-scale Common Watermark Dataset:

• Source: Kaggle - Large-scale Common Watermark Dataset
• Contains realistic visible watermarks for benchmarking removal algorithms.

Quick Start

Install dependencies

pip install -r requirements.txt

Download dataset (optional)

Download the Large-scale Common Watermark Dataset from the link above and place it in datasets/.

Run benchmark (example)

python experiments/benchmark_cnn_only.py \
  --config configs/fast_cnn_only.yaml \
  --input_dir <watermarked_dir> \
  --clean_dir <clean_dir> \
  --wm_suffix v2 \
  --output results/benchmarks/phaseA_cnn_only.csv

• --wm_suffix v2 matches files like xxxv2.jpg (watermarked) to xxx.jpg (clean).

Results

• Per-image metrics and aggregates are saved under results/benchmarks/.

Directory Structure

core/          # Main attack pipeline and CNN detector
utils/         # Image I/O, transforms, metrics
configs/       # YAML configs (e.g., fast_cnn_only.yaml)
experiments/   # Benchmark and evaluation scripts
datasets/      # (Place downloaded datasets here)
requirements.txt
README.md

Reference Performance

Measured on 44 real images from the Large-scale Common Watermark Dataset:

• Success rate: 95.5%
• Avg runtime: 1.5 s / image
• Peak VRAM: 0.26 GB
• Avg PSNR: 23.5 dB

(Results may vary depending on watermark scheme, dataset, and GPU.)

License / Attribution

This repository is intended for research and benchmarking.
For commercial or large-scale use, please refer to and credit the original UnMarker work:

• UnMarker: A Universal Attack on Defensive Image Watermarking
  https://github.com/andrekassis/ai-watermark

Dataset credit:

• Large-scale Common Watermark Dataset https://www.kaggle.com/datasets/kamino/largescale-common-watermark-dataset