License-Plate-Recognition

License plate detection using YOLOv4 trained on custom data.
mAP :- 88.25%(IoU threshold = 50%)
avg IoU :- 62.87%(conf_threshold = 0.25)
avg fps :- 16
License plate text detection and recognition using keras-ocr.
Rejecting false positives by matching pattern with Indian license plates.

Requirements:

   absl-py==0.9.0
   easydict==1.9
   keras_ocr==0.8.3
   opencv-python==4.2.0.34
   tensorflow==2.2.0

Demo:

Download pretrained model from here and copy it inside "data" folder.

 #Run demo on sample video with default arguments
 python demo_video.py
 
 #Run demo with command line arguments 

 python demo_video.py --input "Input_video_path" --output "Path_to_save_result" --size "frame_size"
 
 #Example
 python demo_video.py --input inputs/demo1.mp4 --output results/output1.avi --size 608

#Run demo on sample image with default arguments
python demo.py

#Run demo with command line arguments
python demo_video.py --input "Input_image_path" --output "Path_to_save_result" --size "frame_size"

#Example
python demo.py --input inputs/1.jpg --output results/output1.jpg --size 608

#Note -

Command line argument 'size' must be a multiple of 32. Increasing 'size' increases the accuracy of license plate detection but requires more memory. Reduce the size if your gpu runs out of memory.
If the gpu-ram is 4 GB or less, Reduce memory-limit in this line to a value less than your gpu-ram.
If the gpu-ram is 2 GB or less, Reduce memory-limit here and here to a value less than your gpu-ram.
Since there are 3 models running in a sequence(yolov4 for license plate detection, keras-ocr CRAFT text detection and keras-ocr CRNN text recognition), memory usage is high and fps is low. This solution gives an average fps of 2.5 on gtx 1660 gpu. Fps and memory usage can be improved by training a single YOLOv4 model for both license plate detection and text recognition.

Results:

How to train:

Download dataset zip from this link - https://www.kaggle.com/tustunkok/license-plate-detection/data
Create a folder named 'dataset' inside 'data' and extract the content of zip into 'dataset' folder.
Go to 'data' folder and run this command on cmd - "python augmentation.py". After augmentation total number of images will be 2154.
Now to train darknet YOLOv4 on the dataset, follow the steps given here - https://github.com/AlexeyAB/darknet#how-to-train-to-detect-your-custom-objects

keshavoct98/Indian-Number-Plate-Recognition

License-Plate-Recognition

Requirements:

Demo:

Results:

How to train:

References