deep-anpr
deep-anpr copied to clipboard
Using neural networks to build an automatic number plate recognition system
Deep ANPR
Using neural networks to build an automatic number plate recognition system. See this blog post for an explanation.
Note: This is an experimental project and is incomplete in a number of ways, if you're looking for a practical number plate recognition system this project is not for you. If however you've read the above blog post and wish to tinker with the code, read on. If you're really keen you can tackle some of the enhancements on the Issues page to help make this project more practical. Please comment on the relevant issue if you plan on making an enhancement and we can talk through the potential solution.
Usage is as follows:
-
./extractbgs.py SUN397.tar.gz
: Extract ~3GB of background images from the SUN database intobgs/
. (bgs/
must not already exist.) The tar file (36GB) can be downloaded here. This step may take a while as it will extract 108,634 images. -
./gen.py 1000
: Generate 1000 test set images intest/
. (test/
must not already exist.) This step requiresUKNumberPlate.ttf
to be in thefonts/
directory, which can be downloaded here. -
./train.py
: Train the model. A GPU is recommended for this step. It will take around 100,000 batches to converge. When you're satisfied that the network has learned enough pressCtrl+C
and the process will write the weights toweights.npz
and return. -
./detect.py in.jpg weights.npz out.jpg
: Detect number plates in an image.
The project has the following dependencies:
- TensorFlow
- OpenCV
- NumPy
Different typefaces can be put in fonts/
in order to match different type
faces. With a large enough variety the network will learn to generalize and
will match as yet unseen typefaces. See
#1 for more information.