SAiDL-Spring-Assignment-2022

In this assignment I have attempted the question 3C of the SAiDL spring assignment.

Choosing the data

While training the model with the complete Oxford IIIT pet dataset, the RAM usage was being exceeded hence only 3 pet classes namely Abyssinian, American bulldog, American Pitbull Terrier have been used in the training, validation, testing datasets.

Description

1. Preparing the dataset.

   • Training data - Cropped patches of training images.
   • Testing and Validation data - Random cropped patches of testing and validation images.

2. Building the model

   • Used TensorFlow/Keras to build the SRCNN model.
   • The model used was the classic 9-1-5 model as described in the paper [1501.00092] Image Super-Resolution Using Deep Convolutional Networks i.e. conv layer 1 has filter sizes 9X9, conv layer 2 has filter sizes 1X1, and conv layer 3 has filter sizes 5X5.

3. Training the model with batch size = 64, 20 epochs and loss function being Adam.

4. Plotting the metrics (loss, accuracy) against the epochs.

   • As we see the plots, the loss function reduces and accuracy improves significantly over the epochs.

5. Predicting the super resolution image of a distorted version of an original image

   • Comparing the PSNR scores of Predicted with original image and distorted with original image proves the point that SRCNN improves super resolution of images.