Hough Transform using Convolutional Neural Networks (CNNs)

Given a binary edge-image, this project tries to detect lines using CNNs (Convolutional Neural Networks) by ’switching off’ active pixels that does not belong to any line.

Motivation

Standard HT (Hough Transform) is a popularly used method of estimating lines, given a binary image. However, standard HT requires an accumulator array whose size determines the level of incorporated detail. This introduces a tradeoff between precision and computational cost. Furthermore, the level of detail to be accounted in accumulator array differs from image to image which leads to hyper-parameter optimization for each image.

For more details, please see the project report.

Results

In this repository, we provide

• Dataset generation code
• Training/Testing code to reproduce the results
• Pretrained (only best two) models' weights
• Results of pretrained (only best two) models

1. Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.

1.1. Prerequisites

You need to have following libraries installed:

Skimage >= 0.13.0
Sklearn >= 0.19.1
Numpy >= 1.13.1

Tensorflow >= 1.0.0
Keras >= 2.0.5 
Keras_contrib >= 0.0.2

Pydot
Graphviz

1.2. Installation

1.2.1. Anaconda

Although, packages listed above can be seperately downloaded and installed, it's recommended to install Anaconda package to install all scipy libraries at once.

1. Download Anaconda Installer from here

2. Run the downloaded .sh script with bash: bash Anaconda****.sh

1.2.2. Keras

Use conda package manager to install Keras:

• For CPU version: conda install keras
• For GPU Version: conda install -c anaconda keras-gpu (Note: This will automatically install tensorflow too.)

2. Demo/Quick Start

1. Put the images in directory $dir

2. Go to ./src/ folder

cd src/

2. Run test.py python script. This script will predict images in dataDir and save the results in outDir.

python test.py -dataDir <path-to-test-images> -outDir <path-to-save-results-to> -modelExpName <experiment-log-directory>

(For details on all arguments, please run python test.py --help)

NOTE: Default arguments of test.py are set to run our best model on default data location.

3. Training

3.1. Dataset Preparation

1. Go to ./src/ folder

cd src/

2. Run the generate_dataset.py script to generate the dataset synthetically. This will save input images and corresponding ground truth images at outDir/X/ and outDir/Y/ respectively.

python prepare_dataset.py -outDir <path-to-save-images> -numImgs <number-of-images-to-generate>

(See all arguments using python generate_dataset.py --help)

3.2. Training the model

Run the train.py script to train a model on generated dataset, like so:

python train.py -dataDir <path-to-dataset> -netType <network-name> -logDir <path-to-save-experiment>

(See all arguments using python train.py --help)

This will train a specified model on the specified dataset and will save the following to logRootDir/logDir/:

• Model architecture along with weights
• Tensorboard logs
• Predictions on validation set (of best performing model only)
• Options used for generating this experiment

NOTE: Default arguments of train.py are set to train sequential image-to-image network from scratch.

Author(s)

• Muneeb Aadil
• Sibt ul Hussain