Automated Number Plate Detection is a very important problem in our day to day world, from tolls to security cameras these have enumerous applications. We present you the solution of Mosaic'21 Problem Statement 2, which asks us to make a algorithm which detections the Vehicle Number from the image of its number plate. like so.
- Our algorithm automatically segments License Plate out of the Full image for better accuracy.
- It is robust in all lighting conditions and is not dependent on any hyperparameter of the image.
- Our Algorithm is the most efficient of all methods with such a high accuracy, which enables us to deploy this on entry level hardware easy, making it more accesible.
We have used Yolo v5 to get bounding boxes for characters. We trained it on a custom dataset of 1000 license plates (augmented to 5000 images). We have also used another Yolo v5 to get bounding boxes of license plates from full images as explained in the salient features. Here we used a dataset of 300 images only. Finally for character recognition we had trained a simple CNN model with 36 output layers, trained on CHAR-47 dataset with a bit of data augmentation.
Our architectture is a robust and the most efficient one compared to Unets or Faster RCNN because we have inculcated state-of-the-art Yolov5 models. Below is a flow chart of our approach.
We are able to segmentize using our model in varies lighting conditions and scenarios. We have achieved high accuracy in many Indian and Non-Indian based Number plates. One can find many outputs in "inference/output" folder. Below is the output of our algorithm on a simple yet challenging video.
Note: We have used GTX 1650 to run all demonstrations and achieved a FPS of 20-30.
| Model Name | F1 Score | Giou loss | Validation |
|---|---|---|---|
| License Plate Detector |  |
 |
 |
| Character Segmentation |  |
 |
 |