BINARIZATION

This repository can be packaged into a portable Docker image that:

• installs the runtime dependencies for the orchestrator and both model pipelines
• copies the current checkout during the image build so the code, weights, and SBB model are baked into the image
• can then be run on a GPU-enabled Docker host with bind-mounted input and output directories

Dockerfile

Use the root-level Dockerfile. It copies the current checkout during image build so the code, weights, and model assets are baked into the image. Make sure Git LFS files are present locally before building.

Build

Build the image:

docker build -t binarization-orchestrator:gpu .

Build it from scratch with no cache:

docker build --no-cache -t binarization-orchestrator:gpu .

If you want to build a different ref, check out that ref and run git lfs pull before building.

Run

Run the GPU preflight check first. This loads no model files and does not process images:

docker run --rm --gpus all binarization-orchestrator:gpu --preflight-only

Bind-mount a host input directory and output directory, then pass those mount points to the orchestrator:

docker run --rm --gpus all -v /mnt/PRIMARY/PRIMARY/BINARIZATION/working_slim:/input -v /mnt/PRIMARY/PRIMARY/BINARIZATION/binary_slim:/output binarization-orchestrator:gpu /input /output

docker run --rm --gpus all -v "C:\Users\natha\OneDrive\Desktop\TEST_IMAGE_PROCESSING\working_slim":/input -v "C:\Users\natha\OneDrive\Desktop\TEST_IMAGE_PROCESSING\binary_slim":/output binarization-orchestrator:gpu /input /output --resume

If you want to preserve intermediate pipeline outputs on the host as well:

docker run --rm --gpus all \
  -v /absolute/path/to/input:/input \
  -v /absolute/path/to/output:/output \
  -v /absolute/path/to/work:/work \
  binarization-orchestrator:gpu \
  /input /output --work-dir /work --keep-intermediates

For very large batches or very large source images, keep the SBB subprocess small so TensorFlow memory is released between images:

docker run --rm --gpus all \
  -v /absolute/path/to/input:/input \
  -v /absolute/path/to/output:/output \
  -v /absolute/path/to/work:/work \
  binarization-orchestrator:gpu \
  /input /output --work-dir /work --sbb-chunk-size 1 --resume

The orchestrator processes full-resolution source images through the same DP-LinkNet and SBB scripts; chunking only changes subprocess boundaries and does not resize, downsample, or otherwise alter model inference.

To move the image to another machine:

docker save binarization-orchestrator:gpu -o binarization-orchestrator-gpu.tar
docker load -i binarization-orchestrator-gpu.tar

Input Directory Requirements

The orchestrator expects:

• a flat input directory, not recursive
• at least one image file
• every image filename to start with CROP
• supported image extensions: .png, .jpg, .jpeg, .tif, .tiff, .bmp, .gif, .webp

The output directory can be empty or missing; the orchestrator will create it if needed and will write merged BINARY... TIFF files there. The merge is pixelwise with black winning, so a pixel is black in the final output if either model marks it black.