Hailo's license plate recognition network (lprnet) was trained in-house on a synthetic auto-generated dataset to predict registration numbers of license plates under various weather and lighting conditions.
A convolutional network based on LPRNet, with several modifications:
- A ResNet like backbone with 4 stages, each containing 2 residual blocks
- Several kernel shape changes
- Maximal license plate length of 19 digits
- More details can be found here
- Number of parameters: 7.14M
- GMACS: 18.29
- Accuracy* : 99.96%* Evaluated on internal dataset containing 1178 images
- RGB license plate image with size of 75x300x3
- Image normalization occurs on-chip
- A tensor with size 5x19x11
- Post-processing outputs a tensor with size of 1x19x11
- The 11 channels contain logits scores for 11 classes (10 digits + blank class)
- A Connectionist temporal classification (CTC) greedy decoding outputs the final license plate number prediction
The compiled network can be downloaded from here.
Use the following command to measure model performance on hailo’s HW:
hailortcli run2 set-net lprnet.hef
- Hailo's LPRNet was trained on a synthetic auto-generated dataset containing 4 million license plate images. Auto-generation of synthetic data for training is cheap, allows one to obtain a large annotated dataset easily and can be adapted quickly for other domains
- A notebook for auto-generation of synthetic training data for LPRNet can be found here
- For more details on the training data autogeneration, please see the training guide