Webp Introduction

发表于 更新于 分类于 阅读次数： 本文字数： 4.4k 阅读时长 ≈ 8 分钟

What is Webp?

Webp是一种图像格式，支持单张图片，动图；
什么平台支持？
如何使用？

Open-Source projects

libwebp: https://chromium.googlesource.com/webm/libwebp
mirror: https://github.com/webmproject/libwebp/

User Guide

Options

  1. q: 指定quality factor
  2. alpha_q: 指定透明通道的压缩率
  3. preset: 指定图片的type
  4. -z: 指定lossless压缩的效率
  5. m: 压缩的速度档次，0最快，6是最慢
  6. segment: 指定要使用的segments，默认是4
  7. -size: 指定目标size (in bytes)
  8. -psnr: 指定目标的PSNR
  9. -sns: spatial noise shaping, important
  10. -f: filter strength, default 60，是指loop filter的强度
  11. -sharpness: filter sharpness, 0: most sharp, 7: 最小sharp
  12. -strong: use strong filter
  13. -sharp_yuv: use sharper RGB->YUV的转换
  14. -partition_limit
  15. -pass: analysis pass number
  16. -qrange，指定最小和最大QP的范围
  17. -mt: 使用multi-threading
  18. -alpha_method: 指定透明通道压缩的方法
  19. -alpha_filter: predictive filtering for alpha plane
  20. -exact: preserve RGB value in transparent area; 也就是不修改完全透明区域的
  21. -noalpha: 丢掉alpha的特征
  22. -lossless: 无损压缩图像
  23. -near_lossless: use near-lossless image preprocessing
  24. -hint: 指定图片的特征或者提示；
  25. -af: using autoaf，指标是ssim

如何把其他图片转换成Webp?

JPG format
YUVJ420: YUV420P with JPEG color-range” (i.e. pixels use 0-255 range instead of 16-235 range, where 0 instead of 16 is black, and 255 instead of 235 is white);

使用工具cwebp，最简单的使用就是:
cwebp input.png -q 80 -o output.webp

重要的参数

重要的参数包括-preset -sns -f -m:
-preset: 用于指定content的类型，应该最先被设置，这样后续的设置就在这个基础上来修改; 编码器内部会根据preset去调整sns_strength, filter_sharpness, filter_strength和preprocessing.
-f: 对应的是编码器内部的in-loop filter的filter strength;
-m: 用来指定编码速度，[0, 6]，0最快，6最慢，default是4
-sns: 用来打开一些visual optimization，主要用来调节bit allocation. 用来把easy part的bit分配给hard part。通常升高sns，会导致文件增大，质量变好；

sns

method >= 4: distortion会包括spectural distortion
Distortion: 在method > 4的时候，会有distortion和spectral distortion
设置segment的时候，可以控制幅度。
控制uvac，越小的话，dq_uv_ac变大，和dq_uv_dc，越大,uv_dc boost的越多
It will set do_search

how to calculate alpha:
Do transform, collect coefficients stats, find last_none_zero and max_values;
alpha = last_non_zero/max_value

  1. 计算每个mb的alpha
    原理：
    TX: 4x4
    convert coefficients to bins，calculate the histogram of abs(level)
    maxvalue: 出现系数最多的系数的出现次数，越大，说明0系数的越多，说明能量越少
    last_non_zero: 最后一个非零系数，越大，说明高频的能量越多，
    alpha = ALPHA_SCALE * last_non    +zero / max_value;
    best_alpha = MAX_ALPHA - (3 * best_alpha + best_uv_alpha + 2) >> 2
    高能量:
    1
    2
    3
    4
    5
    MBAnalyze:323 analyze mbx 0, mby 1
    mode 0, last_nonzero_count 31, max_value 95
    mode 1, last_nonzero_count 31, max_value 88
    MBAnalyze:330 best_alpha 179
    MBAnalyze:341 mbx 0, mby 1, alpha 57, uv_alpha 255
    低能量：
    1
    2
    3
    4
    5
    MBAnalyze:323 analyze mbx 0, mby 3
    mode 0, last_nonzero_count 11, max_value 234
    mode 1, last_nonzero_count 4, max_value 244
    MBAnalyze:330 best_alpha 23
    MBAnalyze:341 mbx 0, mby 3, alpha 234, uv_alpha 16
    alpha越大，Q越小；
    beta

filter params

Decoder perspective:
9.4 Loop Filter Type and Levels
frame-level loop filter params:

  1. filter_type: normal or simple, simple filter only applied to luma edges;
  2. loop_filter_level: defines the threshold, if the different is below this threshold, it should be filtered, otherwise, it should not be changed. Usually the level correlated to quantizer levels.
  3. sharpness_level, constant over the frame,
    如果loop_filter_level=0，那么应该跳过loop_filter;
    Differences in excess of a threshold(associated to the loop_filter_level) are assumed to not being modified.

mode_ref_lf_delta_update=1, allow per-macroblock adjustment, through delta; There are two types of delta values, one for reference frame-based adjustment, and the other group is for mode-based adjustment.

Filter header to write into bitstreams:
simple:
level: 对应fstrength_
sharpness

sharpness设置越大，fstrength越大；
Q越大，fstrength越大；

fstrength是针对segment来的；

lf_delta

process;
basestrength = F(q, sharpness)
strength = 5 filter_strength base_strength / *(256 + beta)
level0 = fstrength_

只filter 4x4的块，不filter 16x16和skip的块；

How to decide the strength for each segments?

Why VP8StoreFilterStats? 存储不同的deltaq下，do filter之后的ssim并存储起来；
How to decide final strength;
VP8AdjustFilterStrength

Internal Process

Speed 0/1

Speed 2~6

Alpha Encoding

主要通过一下几个接口来处理

  1. VP8EncInitAlpha
  2. VP8EncStartAlpha
  3. VP8EncFinishAlpha
  4. VP8EncDeleteAlpha
graph TD;
    A-->B;
    A-->C;
    B-->D;
    C-->D;

主要调用EncodeAlpha进行处理

Feature

Specify target_size/target_PSNR

sns:

Region-based quantization

Adaptive Loop-filter

根据prediction mode或参考帧来决定loop filter strength
Adjustment of loop-filter stregnth for each segment.

RDO

Preprocessing

sharp YUV

This is applied when converting RGB to YUV

Smooth segment

pseudo-random dithering log2

Alpha cleanup

Lossless

主要通过调用WebPEncodeLosslessRGB，内部是调用EncodeLossless.
VP8LEncodeImage is used to encoding lossless.
内部调用VP8LEncodeStream
QuantizeLevels
VP8FiltersInit
ApplyFiltersAndEncode
EncodeAlphaInternal —-> EncodeLossless —-> VP8LEncodeStream

Segment

不打开Segment情况下，

rd_opt_level:

  1. no_rd
  2. rd_opt_basic
  3. rd_opt_trellis, perform trellis-quant on the final decision only
  4. rd_opt_trellis_all

if rd_opt_level < rd_opt_basic>, then use VP8EncLoop

Reference

Official-Webp-Doc
React Native using Webp
Webp Tools
VP8 Encode Parameter Guide
VP8 Tech Overview
Webp Compression Techniques