Deepstreet

Deepstreet is the project I developed for my high school thesis in IT @ ITI Marconi, Verona (IT). This project aims to provide a system able to recognize the type of a street sign into an image, using Deep Learning techniques.

Features

• Written in Python
• Keras as a main library for Deep Learning, with Numpy and OpenCV

Getting Started

Prerequisistes

To run all the scripts you need the following packages:

• Python version 3.5
• numpy v. 1.13
• matplotlib v. 2.0
• OpenCV v. 3.0
• tensorflow v. 1.1 -- or tensorflow-gpu if you have GPUs on your machine
• keras v. 2.0
• hdf5 v. 1.8 and h5py v. 2.7

Optional, but recommended:

• NVIDIA cuDNN if you have NVIDIA GPUs on your machine. https://developer.nvidia.com/rdp/cudnn-download

The easiest way to get (most) these is to use an all-in-one installer such as Anaconda from Continuum. This is available for multiple architectures.

Dataset

Download the German Traffic Sign Recognition Benchmark (GTSRB) from https://benchmark.ini.rub.de/

Running scripts

To run the scripts, just download this repo and execute:

python <filename.py>

The two main scripts (deepstreet_training.py and deepstreet_predict.py) can be executed with optional arguments. To find out the arguments for, let's say, deepstreet_training.py execute:

python deepstreet_training.py --help

Using TensorFlow backend.
usage: deepstreet_training.py [-h] [--gpu GPU] [--epochs EPOCHS]
                              [--output_dir OUTPUT_DIR]
                              [--input_dir INPUT_DIR] [--debug DEBUG]

Run a training experiment using pretrained VGG16, specified on the deepstreet
DataSet.

optional arguments:
  -h, --help            show this help message and exit
  --gpu GPU             GPU Device (default: 0)
  --epochs EPOCHS       Number of Epochs during training (default: 10)
  --output_dir OUTPUT_DIR
                        Output directory
  --input_dir INPUT_DIR
                        Input directory
  --debug DEBUG         Debug mode

Contributing Guidelines

All contributions and suggestions are welcome!

For suggested improvements, please create an issue.

License

This project is licensed under the MIT License - see the LICENSE.txt file for details.