Skip to content

Latest commit

 

History

History
80 lines (60 loc) · 4.23 KB

ARCHITECTURE.md

File metadata and controls

80 lines (60 loc) · 4.23 KB

Architecture

This document provides a detailed overview of the technical architecture of the Watermark Remover tool.

Core Components

The system is composed of three main components:

  1. Batch Processor (batch_processor.py): The main entry point and orchestrator.
  2. Complete Watermark Detector (complete_watermark_detector.py): The V3 high-precision detection module.
  3. LaMa Inpainter (lama_inpainter.py): The deep learning inpainting module.
graph TD
    A[User Input] --> B(Batch Processor);
    B --> C{Detection};
    B --> D{Inpainting};
    C --> E[Complete Detector V3];
    D --> F[LaMa Inpainter];
    E --> G[Mask];
    F -- uses --> G;
    F --> H[Clean Image];
    B --> I[Report];
Loading

1. Batch Processor

Responsibilities:

  • Command-Line Interface: Parses user arguments (--input, --output, --batch, etc.).
  • File Discovery: Scans input directories for image files.
  • Orchestration: For each image, it calls the detector and then the inpainter.
  • Reporting: Aggregates results and generates a JSON report and a file list.
  • Error Handling: Catches and logs errors for individual files without halting the entire batch.

2. Complete Watermark Detector (V3)

This is the core innovation of the project, designed to achieve 100% watermark coverage.

Detection Pipeline:

  1. Search Region Identification: Isolates the bottom-right corner of the image to reduce the search space and prevent false positives.
  2. Color-Based Thresholding: Creates an initial binary mask of all pixels matching the watermark's dark gray color profile.
  3. Contour Analysis: Finds all distinct shapes (contours) in the binary mask.
  4. Shape & Size Filtering: Filters contours based on expected width, height, and aspect ratio to identify the primary watermark component.
  5. Three-Stage Coverage Expansion:
    • Method 1 (Direct): Re-analyzes the area within the detected bounding box to ensure all color-matching pixels are included.
    • Method 2 (Expanded): Searches a slightly larger region to capture faint edges.
    • Method 3 (Global): Validates that all detected components are part of the same watermark structure.
  6. Mask Merging: Combines the results of all three methods into a single, comprehensive mask.
  7. Final Refinement: Applies a slight dilation and Gaussian blur to smooth edges and close any microscopic gaps, preparing the mask for the inpainting model.

3. LaMa Inpainter

Responsibilities:

  • Model Loading: Initializes the LaMa model and handles device placement (CPU, CUDA, MPS).
  • Image Preprocessing: Prepares the input image and mask for the model.
  • Inpainting: Executes the LaMa model to fill the masked region.
  • Image Postprocessing: Converts the output tensor back into a standard image format (Pillow Image).

Underlying Library: simple-lama-inpainting

This library provides a high-level wrapper around the official LaMa implementation, simplifying the model download, caching, and execution process.

Data Flow

The data flows through the system as follows:

  1. Input: The user provides an input directory of images.
  2. Processing Loop: The Batch Processor iterates through each image.
  3. Detection: The CompleteWatermarkDetector takes an image path and returns a NumPy array representing the binary mask (255 for watermark, 0 for background).
  4. Inpainting: The LamaInpainter takes the original image path and the generated mask, and returns a Pillow Image object of the clean image.
  5. Output: The Batch Processor saves the clean image to the output directory and records the results.
  6. Reporting: After the loop, the Batch Processor compiles all results into a JSON report.

Design Principles

  • Modularity: Each component is self-contained and can be tested or replaced independently.
  • Robustness: The batch processor is designed to handle errors gracefully, ensuring that one failed image does not stop the entire process.
  • Precision: The V3 detector was built with the single goal of achieving 100% mask coverage to eliminate any post-processing requirements.
  • Usability: The command-line interface is designed to be intuitive, with clear arguments and helpful examples.