VP8视频解码指南

需积分: 42 22 浏览量更新于2024-07-23 收藏 386KB PDF 举报

"VP8 decode guide" VP8解码指南是一个详细介绍VP8压缩视频数据格式及其解码过程的文档。VP8是由On2 Technologies（后被Google收购）开发的一种视频编码标准，它在WebM开源多媒体容器格式中扮演着重要角色，主要用来支持网络上的高清视频流。该文档首先概述了VP8的压缩视频数据格式，这是理解解码过程的基础。VP8的数据格式包括关键帧（I帧）和预测帧（P帧、B帧），其中I帧不依赖于任何其他帧，而P帧和B帧则利用前后帧的信息进行预测编码以提高压缩效率。VP8编码器通过使用块级别的运动补偿和DCT变换来降低图像数据的熵，从而实现压缩。在解码过程部分，文档可能详细阐述了如何解析VP8位流，提取出帧数据，并进行逆变换操作。这涉及到解析宏块布局、解码运动矢量、反量化系数以及逆离散余弦变换（IDCT）。对于每个宏块，解码器会应用相应的运动补偿技术，将预测误差与运动矢量相结合，形成预测帧。然后，解码器对预测误差进行IDCT，恢复图像块的原始像素值。对于B帧，解码器需要处理两个方向的运动补偿。文档还可能讨论了错误检测和恢复策略，因为在网络传输中可能会出现数据丢失或损坏。VP8编码器在编码过程中引入了错误隐藏机制，比如使用独立的编码块和自适应的熵编码，使得即使在某些块受损的情况下，视频的其他部分仍能正常解码。此外，文档可能还涵盖了VP8解码器的实现细节，包括硬件加速支持和内存管理，这对于高效且低延迟的视频播放至关重要。对于开发者来说，这些信息有助于构建高效的VP8解码库或集成到现有多媒体框架中。 VP8虽然不是互联网标准，但它在实际应用中的价值不容忽视。随着互联网视频的广泛应用，理解并正确实现VP8解码对于网络视频服务提供商、浏览器开发商以及嵌入式设备制造商都是至关重要的。 "VP8 Data Format and Decoding Guide"提供了一个全面的指南，帮助读者深入理解VP8压缩格式的内部工作机制，以及如何有效地解码这些压缩数据以重构原始视频帧。这份文档对于开发者和研究者来说是一份宝贵的参考资料，能够促进VP8技术在实际应用中的优化和扩展。

RFC 6386 VP8 Data Format and Decoding Guide November 2011

7. Boolean Entropy Decoder

As discussed in the overview above, essentially the entire VP8 data

stream is encoded using a boolean entropy coder.

An understanding of the bool_decoder is critical to the

implementation of a VP8 decompressor, so we discuss the bool_decoder

in detail. It is easier to comprehend the bool_decoder in

conjunction with the bool_encoder used by the compressor to write the

compressed data partitions.

The bool_encoder encodes (and the bool_decoder decodes) one bool

(zero-or-one boolean value) at a time. Its purpose is to losslessly

compress a sequence of bools for which the probability of their being

zero or one can be well-estimated (via constant or previously coded

information) at the time they are written, using identical

corresponding probabilities at the time they are read.

As the reader is probably aware, if a bool is much more likely to be

zero than one (for instance), it can, on average, be faithfully

encoded using much less than one bit per value. The bool_encoder

exploits this.

In the 1940s, [Shannon] proved that there is a lower bound for the

average datarate of a faithful encoding of a sequence of bools (whose

probability distributions are known and are independent of each

other) and also that there are encoding algorithms that approximate

this lower bound as closely as one wishes.

If we encode a sequence of bools whose probability of being zero is p

(and whose probability of being 1 is 1-p), the lowest possible

datarate per value is

plog(p) + (1-p)log(1-p);

taking the logarithms to the base 1/2 expresses the datarate in bits/

value.

We give two simple examples. At one extreme, if p = 1/2, then log(p)

= log(1-p) = 1, and the lowest possible datarate per bool is 1/2 +

1/2 = 1; that is, we cannot do any better than simply literally

writing out bits. At another extreme, if p is very small, say p =

1/1024, then log(p)=10, log(1-p) is roughly .0014, and the lowest

possible datarate is approximately 10/1024 + .0014, roughly 1/100 of

a bit per bool.

Bankoski, et al. Informational [Page 16]

RFC 6386 VP8 Data Format and Decoding Guide November 2011

Because most of the bools in the VP8 datastream have zero-

probabilities nowhere near 1/2, the compression provided by the

bool_encoder is critical to the performance of VP8.

The boolean coder used by VP8 is a variant of an arithmetic coder.

An excellent discussion of arithmetic coding (and other lossless

compression techniques) can be found in [Bell].

7.1. Underlying Theory of Coding

The basic idea used by the boolean coder is to consider the entire

data stream (either of the partitions in our case) as the binary

expansion of a single number x with 0 <= x < 1. The bits (or bytes)

in x are of course written from high to low order, and if b[j] (B[j])

is the j^(th) bit (byte) in the partition, the value x is simply the

sum (starting with j = 1) of pow(2, -j) * b[j] or pow(256, -j) *

B[j].

Before the first bool is coded, all values of x are possible.

The coding of each bool restricts the possible values of x in

proportion to the probability of what is coded. If p1 is the

probability of the first bool being zero and a zero is coded, the

range of possible values of x is restricted to 0 <= x < p1. If a one

is coded, the range becomes p1 <= x < 1.

The coding continues by repeating the same idea. At every stage,

there is an interval a <= x < b of possible values of x. If p is the

probability of a zero being coded at this stage and a zero is coded,

the interval becomes a <= x < a + (p(b-a)). If a one is coded, the

possible values of x are restricted to a + (p(b-a)) <= x < b.

Assuming that only finitely many values are to be coded, after the

encoder has received the last bool, it can write as its output any

value x that lies in the final interval. VP8 simply writes the left

endpoint of the final interval. Consequently, the output it would

make if encoding were to stop at any time either increases or stays

the same as each bool is encoded.

Decoding parallels encoding. The decoder is presented with the

number x, which has only the initial restriction 0 <= x < 1. To

decode the first bool, the decoder is given the first probability p1.

If x < p1, a zero is decoded; if x >= p1, a one is decoded. In

either case, the new restriction on x -- that is, the interval of

possible values of x -- is remembered.

Bankoski, et al. Informational [Page 17]

RFC 6386 VP8 Data Format and Decoding Guide November 2011

Decoding continues in exactly the same way: If a <= x < b is the

current interval and we are to decode a bool with zero-probability p,

we return a zero if a <= x < a + (p(b-a)) and a one if a + (p(b-a))

<= x < b. In either case, the new restriction is remembered in

preparation for decoding the next bool.

The process outlined above uses real numbers of infinite precision to

express the probabilities and ranges. It is true that, if one could

actualize this process and coded a large number of bools whose

supplied probabilities matched their value distributions, the

datarate achieved would approach the theoretical minimum as the

number of bools encoded increased.

Unfortunately, computers operate at finite precision, and an

approximation to the theoretically perfect process described above is

necessary. Such approximation increases the datarate but, at quite

moderate precision and for a wide variety of data sets, this increase

is negligible.

The only conceptual limitations are, first, that coder probabilities

must be expressed at finite precision and, second, that the decoder

be able to detect each individual modification to the value interval

via examination of a fixed amount of input. As a practical matter,

many of the implementation details stem from the fact that the coder

can function using only a small "window" to incrementally read or

write the arbitrarily precise number x.

7.2. Practical Algorithm Description

VP8’s boolean coder works with 8-bit probabilities p. The range of

such p is 0 <= p <= 255; the actual probability represented by p is

p/256. Also, the coder is designed so that decoding of a bool

requires no more than an 8-bit comparison, and so that the state of

both the encoder and decoder can be easily represented using a small

number of unsigned 16-bit integers.

The details are most easily understood if we first describe the

algorithm using bit-at-a-time input and output. Aside from the

ability to maintain a position in this bitstream and write/read bits,

the encoder also needs the ability to add 1 to the bits already

output; after writing n bits, adding 1 to the existing output is the

same thing as adding pow(2, -n) to x.

Bankoski, et al. Informational [Page 18]

RFC 6386 VP8 Data Format and Decoding Guide November 2011

Together with the bit position, the encoder must maintain two

unsigned 8-bit numbers, which we call "bottom" and "range". Writing

w for the n bits already written and S = pow(2, - n - 8) for the

scale of the current bit position one byte out, we have the following

constraint on all future values v of w (including the final value

v = x):

w + ( S * bottom ) <= v < w + ( S * ( bottom + range ) )

Thus, appending bottom to the already-written bits w gives the left

endpoint of the interval of possible values, appending bottom + range

gives the right endpoint, and range itself (scaled to the current

output position) is the length of the interval.

So that our probabilistic encodings are reasonably accurate, we do

not let range vary by more than a factor of two: It stays within the

bounds 128 <= range <= 255.

The process for encoding a boolean value val whose probability of

being zero is prob / 256 -- and whose probability of being one is

( 256 - prob ) / 256 -- with 1 <= prob <= 255 is as follows.

Using an unsigned 16-bit multiply followed by an unsigned right

shift, we calculate an unsigned 8-bit split value:

split = 1 + (((range - 1) * probability)]] >> 8)

split is approximately ( prob / 256 ) * range and lies within the

bounds 1 <= split <= range - 1. These bounds ensure the correctness

of the decoding procedure described below.

If the incoming boolean val to be encoded is false, we leave the left

interval endpoint bottom alone and reduce range, replacing it by

split. If the incoming val is true, we move up the left endpoint to

bottom + split, propagating any carry to the already-written value w

(this is where we need the ability to add 1 to w), and reduce range

to range - split.

Regardless of the value encoded, range has been reduced and now has

the bounds 1 <= range <= 254. If range < 128, the encoder doubles it

and shifts the high-order bit out of bottom to the output as it also

doubles bottom, repeating this process one bit at a time until 128 <=

range <= 255. Once this is completed, the encoder is ready to accept

another bool, maintaining the constraints described above.

After encoding the last bool, the partition may be completed by

appending bottom to the bitstream.

Bankoski, et al. Informational [Page 19]

RFC 6386 VP8 Data Format and Decoding Guide November 2011

The decoder mimics the state of the encoder. It maintains, together

with an input bit position, two unsigned 8-bit numbers, a range

identical to that maintained by the encoder and a value. Decoding

one bool at a time, the decoder (in effect) tracks the same left

interval endpoint as does the encoder and subtracts it from the

remaining input. Appending the unread portion of the bitstream to

the 8-bit value gives the difference between the actual value encoded

and the known left endpoint.

The decoder is initialized by setting range = 255 and reading the

first 16 input bits into value. The decoder maintains range and

calculates split in exactly the same way as does the encoder.

To decode a bool, it compares value to split; if value < split, the

bool is zero, and range is replaced with split. If value >= split,

the bool is one, range is replaced with range - split, and value is

replaced with value - split.

Again, range is doubled one bit at a time until it is at least 128.

The value is doubled in parallel, shifting a new input bit into the

bottom each time.

Writing Value for value together with the unread input bits and Range

for range extended indefinitely on the right by zeros, the condition

Value < Range is maintained at all times by the decoder. In

particular, the bits shifted out of value as it is doubled are always

zero.

7.3. Actual Implementation

The C code below gives complete implementations of the encoder and

decoder described above. While they are logically identical to the

"bit-at-a-time" versions, they internally buffer a couple of extra

bytes of the bitstream. This allows I/O to be done (more

practically) a byte at a time and drastically reduces the number of

carries the encoder has to propagate into the already-written data.

Another (logically equivalent) implementation may be found in the

reference decoder file bool_decoder.h (Section 20.2).

Bankoski, et al. Informational [Page 20]

剩余303页未读，继续阅读

qq_15687387

粉丝: 0
资源: 1

VP8视频解码指南

vp8_decode.rar

vp8 编解码源代码

【java毕业设计】网页时装购物系统源码（springboot+vue+mysql+说明文档+LW）.zip

Kylin10 + GDAL2.4 + OSG3.6.4 + OsgEarth2.10.1

计算机系统维护技术.xps

数学建模问题中阻滞增长模型

基于Java的菜匣子优选系统设计与实现+jsp（源码）.rar

编程选择题40道：异常处理：错误处理与异常抛出.Tex.docx

为 Vue 2 和 3 创建通用库.zip

LSTM多输入单输出预测

最新资源