Chris Pollett > Students >

    ( Print View)



    [CS 297 Proposal]

    [Deliverable 1]

    [Deliverable 2]

    [Deliverable 3]

    [Deliverable 4]

    [End to End Image Compression - PDF]

    [Image file formats - PDF]

    [DeepN-JPEG - PDF]

    [Image compression using RNN - PDF]

    [CS 297 Report - PDF]

    [CS 298 Proposal]

    [CS 298 report]

Deliverable 3: Modify Deliverable - 1 to run on ImageNet Dataset. Calculate SSIM, PSNR Matrics.


The objective of this deliverable train and test Deliverable 1 on higher resolution images and evaluate the results using matrices such as SSIM and PSNR.

As a part of this deliverable, some experiments were performed on deliverable 1 and some changes were made in the framework.
These experiments are as follows:
  1. Experiment setup: Deliverable -1 was trained on few random images from ImageNet dataset. Since this is a huge dataset, only one subset of this dataset  pets images were considered for training as well as inference.
    Observation: Since images were 200x200 in size as oppose to 32x32 images trained earlier, the program took significantly more time to execute each epoch. Time required for the execution of 1 epoch with 50,000 32x32 images: 53 min
    Time required for the execution of 1 epoch with ~9000 200x200 images: 2 hrs 12 min
    Inference: Significant improvement in the performance of test and train images could be achieved if original images are divided in small parts and trained separately. Since images were stored on Google drive, read speed might have caused the bottleneck.
  2. Experiment setup: Average SSIM and PSNR error was calculated for inference images obtained from Deliverable -1 and Deliverable -4. When few changes were made in the architecture of Deliverable  1. These changes include: Reduction in number of layers in RecCNN from 20 to 13., convolutional kernel size was changed from 3x3 to 2x2. Design strategies used in SqueezeNet were adapted for this experiment.
    Observation: The values are similar.
    Inference: No significant change was observed with these changes.
As a part of this deliverable, a script to calculate image quality metrics (PSNR, SSIM etc.) was written as well. This script requires relative address of two images that needs to be compared and returns the quality scores as well as bit rates of each image as an output.