venkatesh-thiru/SuperRes

clean code of Super Resolution

SuperRes

Clean code of Super Resolution

All the results are obtained with input of 4x4x1.2 compressed images.

The classical interpolation methods we have used are

1. Nearest neighbour
2. Bilinear
3. Bicubic

The result comaparision is as follows

From the results we can see that best result is SSIM .8(median) obtained from bicubic interpollation.

Our implemented DenseNet model gave mean SSIM 0.89(median) results which surpassing all the classical interpolation methods.

Experimenting with differnt kernal sizes:

1. Our first experiment is to compare the performance DenseNet with constant kernal size and varying kernal size. We consider 3X3 kernels for constant kernal size model and 7X7, 5X5, 3X3 kernals for varying kernal size model. We observe that varying kernal size is gave best performance results of SSIM .90(median)

2. We have experimented with different kernel sizes on the DenseNet and found out that starting with a higher size and subsequently reducing the size of the kernel gave us best results. This experiment is done considering 3 different resolution.

L1 loss function for IXI-T1 dataset

The model was trained with the following hyperparameters for IXI-T1 dataset for the L1 loss function: learning_rate = 0.001 Epochs = 50 training_batch_size = 24 validation_batch_size = 6 patch_size = 48 samples_per_volume = 30 max_queue_length = 90

The model was trained for 50 epochs and the output was as follows The training loss was 0.034913947040747316

To remove the grid lines which appeared in the output, following changes were made to the hyperparameters: learning_rate = 0.00001 Epochs = 200 training_batch_size = 24 validation_batch_size = 6 patch_size = 48 samples_per_volume = 40 max_queue_length = 120

The model was trained for 200 epochs and the output was as follows The training loss was 0.0258860532194376

Initial Experiments on IXI-T2 dataset

Loss Function: SSIM

learning_rate = 0.001
Epochs = 50
training_batch_size = 24
validation_batch_size = 6
patch_size = 48
samples_per_volume = 30
max_queue_length = 90

Training Output:

Training loss: 0.9105

Loss Function: L1

learning_rate = 0.00001
Epochs = 200
training_batch_size = 24
validation_batch_size = 6
patch_size = 48
samples_per_volume = 60
max_queue_length = 120

Training Output:

Training loss: 0.021400894038379192