Assignment 4 SOM and Image Data Compression

Neural Networks and Adaptive Systems
(ECE656)
Computer Assignment 4 (SOM and Image Data Compression)
The purpose of this computer assignment is to design a self-organized feature map
(SOFM) to perform vector quantization (VQ) on digital images. The aim is to reduce the
number of bits for efficient representation of images prior to transmission or storage.
There are several images in Computer Assignment 4 folder on Canvas namely Lena and
B that could be used. These images are of size 512x512 pixels with 8 bits per pixel (bpp)
resolution.
1. Partition one of the images used for training into non-overlapping blocks of sizes 4x4
and train various SOFM networks to establish the codebook vectors. To determine
the optimum number of codebook vectors increase the number of neurons in SOFM
from 16 to 256 (e.g., 16, 32, 64, 128, and 256).
2. Determine the topologically ordering of the SOFM that you generated. What is your
observation/conclusion?
3. Devise a scheme to reconstruct the image from the string of codes and using the
codebook vectors. For each case in Part 1, compute the compression rate¥ and
distortion ( 2 2
10 10log dB original error SNR = σ σ where 2 σ original and 2 σ error are variances of
the original and error images, respectively). Present the performance plots, SNRdB
versus bpp, for each case and then experimentally determine the optimum number of
codebook vectors considering both the visual appearance of the reconstructed image
and compression rate versus distortion trade-off.
4. Demonstrate the generalization ability of the optimal SOFM structure on a different
image and comment on the usefulness of the encoding process for real VQ
implementation.
5. Provide a discussion on your results and point out the advantages/disadvantages of
your SOFM network as an image encoding system in a brief report.
¥ In computing the compression rate the number of bits required for block locations and code string for
each coded block should be considered. Here, we assume codebook vectors have already been encoded and
available to the reconstruction system.