Skip to content

Latest commit

 

History

History
212 lines (173 loc) · 7.26 KB

History.md

File metadata and controls

212 lines (173 loc) · 7.26 KB

3.3.0 / 2022-11-21

updates

  • Load ocr params from modelhub configs
  • From v3.3 Support multiple ocr cnn backbone
  • Re-train all OCR models with shufflenet_v2_x2_0 backbone

3.2.0 / 2022-06-09

updates

  • Added brand numberplate detection (examples/ju/inference/detect_brand_np.ipynb)
  • Update auto number grab tools (examples/ju/dataset_tools/auto_number_grab.ipynb)
  • Added fake numberplate detector (examples/ju/train/experimental/froud_numberplate_train.ipynb)

3.1.0 / 2022-03-28

updates

  • Added support for finding 4 number points exclusively within the found bbox
  • Sped up craft postprocessing by cpp bindings
  • Re-train ocr-ua model
  • Re-trained options model
  • Returned to a separate backbone for ocr models
  • Fixed bag with block_cnn in ocr models

3.0.0 / 2022-03-16

updates

  • Refactored code with Sonarqube
  • Added Pipelines
  • Restructured code
  • Added common backbone for ocr models

2.5.0 / 2021-11-24

updates

  • Replaced custom cnn on resnet in option detector model
  • Added fastapi examle

2.4.0 / 2021-11-01

updates

  • Rewrote OCR to PyTorch
  • Restructured project folders and files
  • Added autoloading of datasets and dependent repositories
  • Optimized training options and OCR with PyTorch Lightning
  • Added new dataset tools
  • Updated datasets and models
  • Added experimental feature Orientation Detector
  • Added tensorrt support for OCRs, YOLO and Options Classification models

2.3.0 / 2021-03-11

updates

  • Optimize multiline to one line algorithm
  • Have combined multiline to one line algorithm with nomeroff_net API
  • Added tornado and flask examples

2.1.0 / 2021-03-11

updates

  • Removed is filled or not is filled classification
  • Rewritten options classification on torch
  • Added multiline to one line algorithm
  • Added automatic selection of bevel angle options in np_points_craft.detect
  • Added modelhub module

2.0.0 / 2021-03-01

updates

  • Replaced numberplate segmentation and RectDetector module on object detection(yolov5) and craft
  • Added from_MaskRCNN_datafromat_to_YOLO_dataformat.ipynb dataset convertor
  • Increased the number of examples in the dataset of finding license plate zones
  • Added train example
  • Updated avto-nomer-tool
  • Added ocr eu onnx-convertor
  • Updated demos .py scripts
  • Updated benchmarks .py scripts
  • Fixed all setup*.py needed
  • Fixed all docker files for new requirements needed
  • Updated .html demo
  • Added faster model for finding license plates for the CPU

deprecated

  • DetectronDetector
  • RectDetector
  • MmdetectionDetector

1.0.0 / 2020-08-27

updates

  • Change main version to 1.0.0 beta
  • Updated all examples for new version
  • Fix small bugs in RectDetector
  • Updated all OCR models

0.4.0 / 2020-08-21

updates

  • Updated all code for tensorflow 2.x usage
  • Updated all models for tensorflow 2.x usage
  • Use tensorflow.keras instance keras

deprecated

  • MaskRcnn model cut out
  • tensorflow 1.x not supported now

Centermask2 / 2020-07-15

training

  • Added new cpu ua OCR-model with 'KA' combination

features

  • Added methods that return OCR probabilities get_acc
  • Added newest pytorch Centermask2 model (3x-faster than MaskRcnn)

bugfix

  • fixed 4 points Detector
  • bug with augmented images fixed

0.3.5 / 2019-07-06

model control manager

  • pip3 install nomeroff-net

0.3.3 / 2019-07-06

model control manager

  • Added mcm to nomeroff_net

0.3.1 / 2019-07-06

training

  • Added experimental support for recognition of Georgia (ge) numbers. Recognition Accuracy 97%

features

  • Added latest model autoloader.

0.3.0 / 2019-06-24

training

  • Re-train mask-rcnn model.

bugfix

  • Fix rounding bug in RectDetect

tools

  • Add Mask RCCN dataset tools to auto-nomer-tool

0.2.3 / 2019-05-16

features

  • Added experimental support for recognition of Kazakhstan (kz) 2 line box numbers. Recognition Accuracy 95%.

training

  • Re-train Kazakhstan (kz) numbers recognition model. Get Recognition Accuracy 94%.
  • Re-train options numbers classification model with ["xx_unknown", "eu_ua_2015", "eu_ua_2004", "eu_ua_1995", "eu", "xx_transit", "ru", "kz", "kz_box"] classes output. Get Classification Accuracy 99,9%.
  • Set simplified convolutional network architecture for numberplate classification by default.

0.2.2 / 2019-03-19

features

  • RectDetector: A new perspective distortion correction mechanism has been added, which more accurately positions the number frame. It is activated using the "fixGeometry" parameter, fixGeometry = true
  • Added experimental support for recognition of Kazakhstan (kz) numbers. Recognition Accuracy 91%

training

  • Added a simplified convolutional network architecture for numberplate classification. To train a simplified model, pass the cnn == "simple" to the train method.

bugfix

  • Fixed a critical bug in a RectDetector that could lead to python sticking

0.2.1 / 2019-03-07

features

  • Added CPU and GPU docker files.
  • Added ru region detection in license plate classification.
  • Added ocr russian number plate detector.

training

  • Update augmentation(use module imgaug).
  • Added freeze model graph and use .pb models in prediction.

0.2.0 / 2019-02-21

features

  • OCR: GRU-network trained on Ukrainian and European license plates are used instead of tesseract).
  • Implemented batch processing of multiple images.
  • The license plate classification model has been improved. Now, a single pass classification has become possible according to different criteria: by type of the license plate and by characteristic are painted / not painted.

optimizations

  • Implemented asynchronous versions of the set of methods, which gives a performance increase of up to 10%.
  • Optimized code for use on Nvidia GPUs.

training

0.1.1 / 2019-01-17

features