This is a PyTorch implementation of a Deep Convolutional Neural Network model for detecting the parameters of a circle presents inside a given image under the presence of noise.

Note: This implementation is using Multi-GPUs for faster training. You can modify train.py to avoid using GPUs.

The output of the network is 3 real numbers which represents the detected row, column, and radious of the circle in the noisy image.

Installation

• Clone this repo:

git clone https://github.com/hsouri/Cirlcle-Detection-CNN

• Requirements Install the dependencies by running the following command:

pip install -r requirements.txt

• Data set making Before trainig, train set should be created by the following command:

python dataset.py

This will generate 200,000 images with randon level of noise between 0.035 and 3.5. You can create yout own train set with arbitrary number of imaages and arbitrary level of noise by changing number of images and level of noise in the train_set() function.

• Training:

python train.py

You can use your own data set by changing the default dataset by the following command:

python train.py --data {directory path to your csv train set}

• Resume training from a saved model:

You are able to resume training your model from a saved checkpoint by running the following:

python train.py --resume {directory path to your saved model}

You can also change other attributes such as batch size, learning rate, number of epochs, number of workers, resume and continue training from a checkpoint. List of selectable attributes:

'--name', '--out_file', '--workers', '--batch-size', '--resume', '--data', '--print_freq', '--epochs', '--start_epoch', '--save_freq'

You can test your model on 1000 random images with level noise of 2 by running the fowwloing command:

python validation.py

By default validation is using the my pretrained model. If you're willing to test your own model you must change the checkpoint in the find_circle() function.

Losses for trainig the model will be saved in the out.txt file. Trinig loss of 50 epochs is as follows:

Validation scores using the metric AP 0.7 are as follows: