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
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)
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
updates
- Refactored code with Sonarqube
- Added Pipelines
- Restructured code
- Added common backbone for ocr models
updates
- Replaced custom cnn on resnet in option detector model
- Added fastapi examle
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
updates
- Optimize multiline to one line algorithm
- Have combined multiline to one line algorithm with nomeroff_net API
- Added tornado and flask examples
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
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
updates
- Change main version to 1.0.0 beta
- Updated all examples for new version
- Fix small bugs in RectDetector
- Updated all OCR models
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
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
model control manager
- pip3 install nomeroff-net
model control manager
- Added mcm to nomeroff_net
training
- Added experimental support for recognition of Georgia (ge) numbers. Recognition Accuracy 97%
features
- Added latest model autoloader.
training
- Re-train mask-rcnn model.
bugfix
- Fix rounding bug in RectDetect
tools
- Add Mask RCCN dataset tools to auto-nomer-tool
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.
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
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.
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
- A small nodejs admin panel was created, with which you can prepare your dataset for license plate classification or OCR text detection tasks.
- Prepare example script for OCR train.
- Prepare example script for Options Classification.
- Added numberplate MaskRCNN example script.
features
- Add online demo numberplate recognition https://nomeroff.net.ua/onlinedemo.html