Report (April 99 to June 99) on research project funded by SAS
 
  1. CA Based Vector Quantization Scheme
We have developed an algorithm for fast VQ of images. The motivation for this stems from the fact that VQ has very low decoding time. However VQ suffers from the disadvantage of high encoding time due to the search procedure. We have worked on this and have arrived at a solution which can give us the target code book index in considerably lesser time - time reduced by more than two order of magnitude. The application area which we are targeting is hand held video phones where essentially portrait (i.e. Lena,claire etc) type of images will be transmitted. We have used a training set of 12 images(portraits). Using this we have determined that there can be about X number of distinct ( our definition) code book entries. Around this X number of code book entries we grow clusters of likely blocks from our training sets. With this input Set we build a CA based classifier which when presented with a target block identifies the the code book entry by which the block can be replaced in almost constant time. We have used X=8192 which implies that we are using 13 bits to encode a 8 X 8 block giving us an effective bpp of 13/(8x8x8) = 0.2539. The psnr of images presented to the encoder is around the 25db mark. This method requires the storing of 8192 code book entries on the decoding side and a similar number of CA configuration entries on the encoding side. This is about 4 MB of memory on either side. Given this, the encoder has a very fast search response time. Executed in software, this method has shown very good response during encoding. If we can use hardware, the speed up will be significantly high. The CA based classifier is the core of our scheme and this essentially reduces the search time by a significant amount. A few sample results is presented in this table. The image size is 352 x 240. Also a sample original and reconstructed image is attached embedded.
 
 
Image PSNR
Julie 33.53
Girl1256 34.58
Michelle 32.69
Girl256 27.84
Claire256 29.91
ash 27.84

 
 
 
 
 
  1. A high speed low cost cellular architecture for Wavelet Transform
We are also working on a new high speed low cost, cellular architecture implementing the Wavelet Transform. The basic methodology is noted below.

With Wavelet transform coefficients as C0,C1,C2,C3……………..Ck . The transformed pixel value X’ ( for the original value X ) is conventionally derived through multiplication of current pixel value with the coefficient and next add operations. The proposed architecture eliminates multiplication and employs CSA( Carry Save Adder )while achieving same order of PSNR value. Dedicated architecture achieving higher speed of operation with lesser cost is the targeted motivation. In addition we are also looking into the possibility of reducing the ring effect usually observed with the wavelet scheme.