The data was generated using java/DistortImage/src/image/DistortImage.java. This process downloaded images from ImageNet, sized down the image to 224x224 pixels, and applied a series of six distortions on each image.

Original Image:

Gaussian (Smooth) Blur:

Motion Blur:

Non-Monochrome Gaussian Noise:

Monochrome Gaussian Noise:

Marble:

Twirl:

We trained a softmax regression model as a baseline using softmaxRegression/softmax.py.

We trained VGG16 model from scratch using vgg16/vgg16_baseline.py.
Scripts for evaluation of the VGG16 models are in 'vgg16'.

We fine tuned VGG16 model using transferLearningVGG16/transfer_vgg16.py.
Scripts for evaluation and error analysis of the VGG16 models are in 'transferLearningVGG16'.

We fine tuned SGDNet model using sgdnet_transfer/transfer_sgdnet.py. Scripts for evaluation and error analysis of the SGDNet transferred models are in 'sgdnet_transfer'.

We calculated structural similarity index metric using ssim.py .
Scripts for evaluation of VGG16 and SGDNet on test sets using ssim are in 'ssim_calc.ipynb'