Technical test @ ParkDepot

A deep learning classifier on German license plates 🚘 🇩🇪

Task

• Detect if a plate has 1, 2 or 3 letters for the city description :
• Implement a deep learning classifier

Dataset

• Real plates
• Synthetic plates, taht have been generated to construct a bigger dataset
• Input, example of such a plate:

Used method

• A strong preprocessing pipeline to split data into usable training/validation sets and to transform data dimension
• A manually implemented ResNet model, following the original paper
  • Redefinition of the residual blocks and stacks in PyTorch
• Use of a classic SGD optimizer (lr=1e-2, momentum=0.9, weight_decay=1e-2)
• Use of Google GPU capacity, hence the Google Colab notebook.

Results

• High final accuracy (0.999%) on test data