Vehicle-Number-Plate-Detection

This project demonstrates the use of TensorFlow Object Detection API to automatically number plates (Indian) from vehicles.

Dataset used: https://www.kaggle.com/dataturks/vehicle-number-plate-detection

File description:

• Data-Images.zip: Contains the images of the cars, number plates and annotations in .txt files (YOLO format)
• Data_prep_and_visualization.ipynb: A notebook demonstrating the process of preparing the dataset (.csv files) for creating TFRecords (otherwise TensorFlow Object Detection API won't work)
• Indian_Number_plates.json: Configuration file which contains image download paths and annotations
• exported_graph: Contains the inference graph in .pb and .tflite formats which can be used to run inference on both CPU platforms and on-device platforms
• label_map.pbtxt: Contains the encodings of the dataset classes which,in this case, is 1: license_plate
• ssd_mobilenet_v1_pets.config: Training and evaluation pipeline configuration file as needed by TensorFlow Object Detection API
• test.record & train.record: TFRecords files of testing and training sets respectively
• test_labels.csv & train_labels.csv: .csv files as required by the generate_tfrecord.py script

Demo inference (as collected from TensorBoard):

To kick-start the model training process and to export the trained model's inference graph (using the export_tflite_ssd_graph.py script), I followed:

• TensorFlow Object Detection API's official documentation
• Training and serving a realtime mobile object detector in 30 minutes with Cloud TPUs

I used SSD_MobileNet_V1 architecture which was pretrained on the COCO dataset.

To convert the frozen inference graph, I ran the following command:

tflite_convert \
    --output_file=detect.tflite \
    --graph_def_file=frozen_inference_graph.pb \
    --input_shapes=1,300,300,3 \
    --input_arrays=normalized_input_image_tensor \
    --output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3'  \
    --inference_type=QUANTIZED_UINT8 \
    --mean_values=128 \
    --std_dev_values=128 \
    --change_concat_input_ranges=false \
    --allow_custom_ops

Note: To be able convert an inference graph to its .tflite variant you need to enable quantization aware training and you can specify that in the .config file itself.