ImageDT

Python 3 utilities for image processing, files, annotations, metrics, synthetic text, TFRecords and model conversion. The original module paths, function names and public argument names are retained, including historical spellings such as noise_padd, fillter_file_type, wirte_xml_file and gernerate_infos_dict.

Install

Requires Python 3.10 or newer. The candidate release is 0.6.0rc1.

python -m pip install imagedt
# Optional capabilities:
python -m pip install 'imagedt[plot,excel,pdf]'
python -m pip install 'imagedt[tensorflow]'  # TensorFlow 2 + LiteRT
python -m pip install 'imagedt[lite]'        # LiteRT inference without TensorFlow
python -m pip install 'imagedt[caffe]'       # Caffe protobuf/graph tooling

For this checkout, use python -m pip install -e '.[dev,plot,excel,pdf,tensorflow]'. The candidate is only available on PyPI after publication. OpenCV's headless package supports processing without requiring a desktop display; cv_show needs a GUI-enabled OpenCV installation. Install exactly one OpenCV distribution in an environment.

Importing imagedt does not import TensorFlow, matplotlib, OpenCV or spreadsheet backends. Capabilities load when their exported names are accessed. An unavailable optional backend reports the extra to install.

Familiar interfaces

import imagedt

paths = imagedt.loop('./images', ['.jpg', '.png'])
imagedt.write_csv([['name', 'count'], ['图片', 3]], 'counts.csv')
boxes = imagedt.Anno_OP.parse_lab_xml('annotations/example.xml')

FilesOp, Anno_OP, RecordWriter and plot_tools remain singleton objects at their historical import paths. Traversal is deterministic. Text files use UTF-8; CSV readers also accept a UTF-8 BOM.

from imagedt.tensorflow.tools.tfmodel_wrapper import TFmodel_Wrapper

with TFmodel_Wrapper('model.pb', input_nodename='data',
                     output_nodename='softmax,features') as model:
    classifications = model.predict(images)
    features = model.extract(images)

Graph wrappers own isolated sessions and support close() and context managers. Importing a backend does not disable eager execution or change GPU environment variables. TF1 graph APIs use tf.compat.v1; queue-based Data_Provider should be created within tf.Graph().as_default() and run with a Session and queue-runner lifecycle.

Behavioral corrections and migration limits

See the migration guide for precise changes. Important corrections include class-aware detection matching, filename-aligned classification evaluation, no train/validation leakage, bounded dataset batching, binary XML writing, complete image cleanup, and collision-aware paired renaming.

Maintained conversion support covers the bundled Caffe layer subset to frozen GraphDef, frozen GraphDef to float TFLite, and float32 single-image NHWC LiteRT classification. Arbitrary Caffe variants or model checkpoints are not guaranteed. LeNet uses Keras layers while retaining endpoint keys; old TF-Slim variable/checkpoint names are not preserved.

These old capabilities retain callable entrances and raise explicit errors:

• Caffe-to-CoreML: removed from modern coremltools; first migrate the model to a supported format.
• Handwriting generation: the external TF1 pretrained model was never bundled. Computer text generation remains supported.
• Legacy TOCO GRAPHVIZ_DOT and uint8 quantization without representative calibration data.

PDF rendering uses portable PDFium wheels rather than ImageMagick/Ghostscript. Single-page output keeps the requested filename; multiple pages use <stem>-0, <stem>-1, etc.

Contributing

See CONTRIBUTING.md for Python conventions, branches and contribution workflow; TESTING.md for checks and test lanes; SECURITY.md for private vulnerability reporting.

Develop and verify

python -m pip install -e '.[dev,plot,excel,pdf,tensorflow]'
python -m ruff check imagedt tests scripts test
python -m ruff format --check imagedt tests scripts test setup.py
python -m pytest --cov=imagedt --cov-report=term-missing
python -m build
python -m twine check dist/*
python scripts/benchmark_utils.py

Tests generate their own images, XML, fonts and tiny models in temporary directories. No personal /data paths, production datasets or GPU are needed. TensorFlow tests skip explicitly when the optional backend is absent. tests/fixtures/legacy_api.json captures public names and argument names from the original revision and guards compatibility. It does not claim numerical equivalence for intentionally corrected bugs or unchanged trained-checkpoint compatibility.

CI defines Python 3.10/3.12/3.14 core jobs on Linux, macOS and Windows, plus a Python 3.12 TensorFlow/LiteRT integration job. Remote CI execution is separate from local verification.

Release

bash upload_pip.sh builds into a fresh temporary directory and checks metadata. It does not upload. Publish only after reviewing the artifacts and configuring PyPI credentials:

PYTHON=python3 bash upload_pip.sh --upload

See release validation and changes. License: Apache-2.0.

The historical technical review records the original revision baseline, architecture and migration decisions.