Table of Contents

• Introduction
• Backgound and Problem Statement
  • ZRLE
• The based theory
  • YUV-RGB color transform
  • Wavelet transform
  • Important observations of wavelet coefficients
• ZYWRLE Definition
  • Specification about Wavelet coefficients filtering
  • Specification about Wavelet transform
  • Changes From ZRLE
• Evalution of ZYWRLE
  • Compression performance in motion picture
  • Errors in YUV conversion
• Implemantation of ZYWRLE
  • Our VNC client for embeded device:"micro-VNC"
  • Reference Implementation(C source code/patch for VNC Servers)
• License of ZYWRLE
• To Do

The lossy codec based on Wavelet transform is frequently used in compression of natural image, motion picture and audio stream.
But it isn't almost used in Remote Desktop protocol like VNC because classical lossless run-length codec like ZRLE has high performance in peculiar image which has big mono color region like desktop image.
The lack of lossy codec cause low performance of motion picture in VNC.
We introduce new codec "ZYWRLE(ZLib YUV Wavelet Run Length Encoding)" which applied wavelet transform partially and is based on ZRLE.
In this new codec, network traffic reduced to 15%-50% traffic of ZRLE when we play motion picture.

Backgound and Problem Statement

ZRLE : See "RFB Protocol Specification".

Problem of ZRLE:

• In Hi-color(above 32K colors) mode,
• In Natural image,

For solving problem of ZRLE for motion picture:
=>Reduce Raw tile.

• Decrease using color
• Make many "Run"
• Bias pixel probability

The based theory

• Natural images in general have a low pass spectrum. When an image is wavelet transformed the energy in the subbands decreases as the scale decreases (low scale means high resolution), so the wavelet coefficients will, on average, be smaller in the higher subbands than in the lower subbands. This shows that progressive encoding is a very natural choice for compressing wavelet transformed images, since the higher subbands only add detail.(From EZW.)
• Large wavelet coefficients are more important than small wavelet coefficients. (From EZW.)
• Wavelet coefficients distribute around 0 according to the bell-shaped distribution like Gaussian. Lets denote the standard deviation of the distribution as σ. By discarding small wavelet coefficients, the amount of the data can be reduced. (From Image Compression using the Wavelet Decomposition.)

ZYWRLE Definition

• We use Non liner quantization filter.

  y=f(x) = sign(x)*round( ((abs(x)/(2^7))^ r )* 2^(o-1) )*2^(8-o)
  x=f-1(y) = sign(y)*round( ((abs(y)/(2^7))^(1/r))* 2^(i-1) )*2^(8-i)
  ( r:power coefficient i:effective MSB in input o:effective MSB in output )

  r < 1.0 : Smaller value is more important than larger value.
  r > 1.0 : Larger value is more important than smaller value.
  r = 1.0 : Liner quantization which is same with EZW style.

  r = 0.75 is famous non liner quantization used in MP3 audio codec. In contrast to audio data, larger value is important in wavelet coefficients. So, I select r = 2.0 table( quantize is x^2, dequantize sqrt(x) ).

  Filter specification(r = 2.0)

  quality	Wavelet level	Y Filter	UV filter
  Hi	1	o=4	o=0
  Mid	2	o=2,5	o=0,5

• For PLHaar, Pixel range change to "Odd range(-127<=x<=127)". -128 is converted to -127.
• I used 'interleave' decomposition instead of the famous 'pyramid' decomposition for more speed and less memory.
• 4 sub coefficient Wavelet transform image decompositon Hxy->Hy->Hx->L for apply ZRLE at once.

• Only Raw tile is applied Wavelet transform. Other tile(RLE,Packed pixel,Palette) is encoded with original image.
• Raw tile threshold was changed to [w]*[h]*[pixel byte]/(2^level) instead of [w]*[h]*[pixel byte].
• In Raw tile, whole tile image is converted with RCT. and next Wavelet transform.
• Wavelet transformed raw tile was encoded ZRLE recursively.

Evalution of ZYWRLE

• Comp. rate of ZRLE(about 30%) goes to 15%(Hi-quality)/5%(Mid-quality)!
• Low-freq. coefficients is "half(level1)/quarter(level2) reduced resolution" original image.
  So, these must be Raw tile and these size is 1/(2*2)=25%, 1/(4*4)=6.25%.
• Remaining Hi-freq. coefficients images go to:
  • level1: 75% => 32%-25%=7%(about 10 times!)
  • level2: 93.75% => 12.1%-6.25%=5.85%(about 20 times!)
  These indicates that Wavlet transform is effective for ZRLE codec with natural image.

There are two type errors in YUV conversion.

• For keeping integer range, UV component reduces to half range.( 0 <= Y <= 255, -255 <= UV <= 255 in original RCT.)
• For PLHaar, Pixel range change to "Odd range(-127<=x<=127)". -128 is converted to -127.

For these reasons, YUV conversion includes errors and PSNR indicates bad quality though there are no distinguishable image distortion.

This error is often distinguishable at border between mono color ZRLE block and ZYWRLE block.

We need more enhancement about YUV conversion.

Implemantation of ZYWRLE

• Core Codec of ZYWRLE(zywrletemplate.c:VNC Server/Client independent)
• Patch for Ultr@VNC V1.0.2(UltraVNC-102-ZYWRLE-diff.txt:include zywrletemplate.c)
• Patch for LibVNCServer 0.8.2(LibVNCServer-0.8.2-ZYWRLE-diff.txt:include zywrletemplate.c)
• Patch for RealVNC/Free for Win V4.1.2(vnc-4_1_2-winsrc-ZYWRLE-diff.txt:include zywrletemplate.c)

• Research whether we should use YUV conversion or not.
  (Is all UV cutting of level 1 too much less quality?)
• Research whether we should use PLHaar or not.
  (Selecting Daubechies or 3-5 Lifting may be better.)
• Research best coefficients filter.
  (Should we use Non-linar quantizer or not? Which is best param of it?)