- Tensorflow
- Numpy
- cv2
- imutils
You can run the demo by running "python3 finalPrototype.py"
In Yolo training folder, there are some cfg file, weights, python code we used to train our 2 yolos
In CNN training folder, there is the python code we used to train our CNN for character recognition
You can donwload pb files, yolo.weights and datasets here : https://drive.google.com/drive/folders/17gxw7tv7jy3KgJFhQiHX0IilYObFbIJp?usp=sharing
More informations : https://medium.com/@theophilebuyssens/license-plate-recognition-using-opencv-yolo-and-keras-f5bfe03afc65
### Criação de ambiente conda sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))#criar ambiente python 3.7.10
conda create --prefix /media/jones/datarec/lpr/env python=3.7.10
#ativar ambiente
conda activate /media/jones/datarec/lpr/env
#instalar tensorflow
pip install tensorflow-gpu==1.14.0
#instalar keras
pip install keras==2.2.4
pip install opencv-python==3.4.2.17
pip install opencv-contrib-python==3.4.2.17
pip install Cython --install-option="--no-cython-compile"
#ir para a pasta do darkflow pra instala-lo:
####TheophileBuy#2
####Just build the Cython extensions in place. NOTE: If installing this way you will have to use ./flow in the cloned darkflow directory instead of flow as darkflow is not installed globally.
python3 setup.py build_ext --inplace
####Let pip install darkflow globally in dev mode (still globally accessible, but changes to the code immediately take effect)
pip install -e .
Install with pip globally
pip install .
###cd /media/jones/datarec/lpr/fontes/ocr/darkflow/darkflow APENAS COMENTADO, NAO UTILIZAR
pip install .
pip install imutils
pip install -Iv h5py==2.10.0
pip install --upgrade Pillow
problema ao definir o LR. encontrada issue no proprio repo:
Darkflow does not use the learning rate in .cfg. Use --lr instead.
https://github.com/thtrieu/darkflow/issues/515#issuecomment-356474112
- Transformar os arquivos de ckp para o formator protobuf (.pb)
https://github.com/thtrieu/darkflow/issues/869#issuecomment-412757194
python flow.py
--model
cfg/yolo-character.cfg
--load
-1
--savepb
- Definir o parâmetro load no train-character.py (utiliza o ultimo checkpoint):
"load": -1
- DS-CHAR-1_train: 1.359 plates
- DS-CHAR-2_train: 4.731 plates
- DS-CHAR-2_test: 1.242 plates / 8.694 chars
- DS-CHAR-3_train: 7.283 plates
- DS-CHAR-3_test: 1.816 plates / ??? chars
- DS-CHAR-RECOG-1_train: 22958 chars
- DS-CHAR-RECOG-1_test: 5740 chars
- DS-CHAR-RECOG-1_train: 22958 chars
- DS-CHAR-RECOG-1_test: 5740 chars
- Letra I agrupada com o numero 1 e letra O agrupada com numero 0
- model size: width=608 height=608
- learning_rate=0.001
- burn_in=2000
- max_batches = 500200
- policy=steps
- steps=10000,20000,30000
- scales=.1,.1,.1
- batch = 8
- epoch = 200
- moving ave loss ~= 6.4
- train DS-CHAR-1_train
- model size: width=416 height=416
- learning_rate=0.001
- burn_in=2000
- max_batches = 500200
- policy=steps
- steps=10000,20000,30000
- scales=.1,.1,.1
- batch = 8
- epoch = 200
- moving ave loss = 3.683196278051683
- train DS-CHAR-1_train
- model size: width=416 height=416
- learning_rate=0.001
- burn_in=5000
- max_batches = 500200
- policy=steps
- steps=10000,20000
- scales=.1,.1
- batch = 8
- epoch = 100
- moving ave loss = 6.2
- train DS-CHAR-2_train
- model size: width=320 height=320
- learning_rate=0.001
- burn_in=2000
- max_batches = 500200
- policy=steps
- steps=20000,50000, 100000
- scales=.1,.1,.9
- batch = 16
- epoch = 120
- moving ave loss = 4
- train DS-CHAR-2_train
- Total de amostras: 8694 chars
- IOU = 0.5
- DS-CHAR-2_test
| Modelo | Precision | Recall | True Positive | False Positive | False Negative |
|---|---|---|---|---|---|
| Baseline Yolov2 | 0.90 | 0.41 | 3522 | 269 | 5172 |
| Yolov2-ceia-v3 | 0.98 | 0.99 | 8597 | 149 | 97 |
| Yolov2-ceia-v4 | 0.99 | 0.99 | 8596 | 130 | 98 |
| Yolov2-ceia-v6 | 0.99 | 0.99 | 8634 | 103 | 60 |
| Yolov2-ceia-v7 | 0.98 | 0.99 | 8624 | 147 | 70 |
| Yolov2-ceia-v8 | 0.99 | 0.99 | 8642 | 93 | 52 |
- DS-CHAR-RECOG-1_train
- model size: width=42 height=63
- epoch = 180
- loss = 0.05
- train DS-CHAR-RECOG-1_train
- model size: width=42 height=63
- epoch = 180
- loss = 0.05
- 2 layers dropout 0.5
- train DS-CHAR-RECOG-1_train
- model size: width=42 height=63
- epoch = 146
- best epoch = 102
- train DS-CHAR-RECOG-1_train
- ResNet20 v1
- model size: width=42 height=63
- epoch = 300
- loss = 0.05
- 2 layers dropout 0.5
- train DS-CHAR-RECOG-2_train
- model size: width=42 height=63
- epoch = 200
- best epoch = 102
- train DS-CHAR-RECOG-2_train
- model size: width=42 height=63
- epoch = 200
- best epoch = 102
- train DS-CHAR-RECOG-2_train
- data augmentation = True
- model size: width=42 height=63
- epoch = 200
- best epoch = 102
- train DS-CHAR-RECOG-2_train
- data augmentation = True
- zoom_range=[0.7, 1.3]
- data augmentation apenas com o gaussian blur.
- houve melhora significatgiva no desempenho
- melhoras: treinar em conjungo com variacao de zoom, shift e brilho
- model size: width=42 height=63
- epoch = 200
- best epoch = 88
- train DS4-CHAR-RECOG-2_train
- data augmentation = True
- gaussian blur
- Total de amostras: 8694 chars
- IOU = 0.5
- DS-CHAR-2_test
| Modelo | loss | acc |
|---|---|---|
| char_recog_ceia_1 | 2.67 | 0.8517 |
| char_recog_ceia_2 | 0.9905 | 0.9024 |
| Ceia_ResNet20v1_model | 0.5065 | 0.93399 |
| char_recog_ceia_3 | 0.85 | 0.9123 |
| ceia_char_recog_2_ResNet20v1_model.086 | 0.4451 | 0.94283 |
| ceia_char_recog_3_ResNet29v2_model | 0.4534 | 0.9412 |
| ceia_char_recog_4_ResNet29v2_model | 0.3613 | 0.9409 |
| ceia_char_recog_5_ResNet29v2_model | 0.3486 | 0.9442 |
| ceia_char_recog_7_gaussian_ResNet29v2_model.088.h5 | 0.29 | 0.9541 |