Pre-Transform Scanning for Video Compression

Abstract

Most video compression algorithms today use block based coding. An image or video frame data is coded one block at a time. A block of data (or residual data block in inter frames) is de-correlated using transform coding to exploit spatial redundancies in the data blocks. Transforms such as 2D DCT are used to de-correlate the video/image data to improve compression performance. The transforms used have a strong energy compaction property; after a transform is applied, most of the signal information is concentrated in a few low-frequency coefficients and results in fewer non-zero coefficients. Fewer non-zero coefficients lead to a lower bitrate and higher compression. In this thesis we have explored novel ways of improving the correlation of the data in a block in order to reduce the number of non-zero coefficients. The number of non-zero coefficients in the transformed data decrease with the increase in spatial correlation in the input data. Similarities among neighboring blocks are exploited by permuting block data using the sorting permutation of the neighbors to improve the correlation. We show that this permutation based approach reduces the number of non-zero coefficients significantly and improves compression. Performance of this method is demonstrated in an implementation in Xvid MPEG-4 video encoder.

From the results we were able to come to certain conclusions regarding the Peak Signal to Noise Ratio (PSNR) and the average number of non zero coefficients (NZC) after transformation. We were able to decrease the average NZC per frame but with loss in PSNR. Consistency was maintained as the approach was extended and implemented for various approaches and also implementation in the Xvid MPEG-4 encoder. Here in the Xvid Mpeg-4 encoder implementation we also concluded that we need a different Variable length Coder (VLC) that supports this approach as the final bitstream resulting out of the VLC was increasing in size.

From the study of the various results from all the different approaches we came to a conclusion that we need to have a new VLC for this approach as concluded from the Xvid implementation results. The proposed approach can also be extended and explored as a post transform scanning approach to see if this approach can be made to loose less PSNR.