Enhanced Method and Apparatus for Image Compression

Abstract

Provides is a method of image compression and decompression involving a variable-to-variable coding scheme, for example as used or of similar nature to a HT Cleanup encoding algorithm, as described in ISO/IEC 15444-15:2019. The method involves a lookup table that the encoder addresses using a potentially modified version of the EMB pattern, a potentially modified version of the significance pattern, and any additional context information, in order to find the associated variable length codeword. The modified significance pattern and modified EMB pattern, each consisting of 1 bit per sample in a quad, implicitly identify the quads original significance pattern and the quads 1-bit unsigned residual offset value, together with some or all of the quads original EMB pattern, both when the unsigned residual offset value is 1 and, for specific values of the significance pattern, when the unsigned residual offset value is 0.

Description

FIELD OF THE INVENTION

The present invention relates to a method and apparatus for image compression, and particularly, but not exclusively, to an enhancement to the HT block coding method described in ISO/IEC 15444-15:2019, known as High Throughput JPEG 2000.

BACKGROUND OF THE INVENTION

The High Throughput JPEG 2000 standard documented in ISO/IEC 15444-15:2016 describes a block coding algorithm consisting of Cleanup, SigProp and MagRef coding passes, of which the most distinctive feature is the HT Cleanup pass. The applicant's earlier patent applications, published as WO2017/201574, WO2019/191811 and WO2020/073098 also provide explanation of the high throughput block encoding technology. These documents provide explanation of the applicant's earlier technology for which the present document discloses enhancements.

The HT Cleanup pass involves three bit-streams, known as the MagSgn, VLC and MEL bit-streams, that collectively encode a rectangular block of sample values of width Wblk and height Hblk, where samples indexed by n have sign s_n and magnitude μ_n. Samples are encoded or decoded in 2×2 blocks, and samples are traversed in a quad-based scanning pattern, such that consecutive indices n address the top-left, bottom-left, top-right and bottom-right samples of each 2×2 quad before advancing to the next quad within a pair of lines, and then on to subsequent line-pairs.

Each quad q is assigned a 4-bit significance pattern p_q that indicates the significance a σ_n of each sample n in the quad, where a σ_n=1 means that the sample is significant, which is true if and only if μ_n#0, while σ_n=0 means that the sample is insignificant. Each sample is represented indirectly via its MagSgn value v_n, together with its significance σ_n, where

${\nu }_n=\{\begin{array}{cc}2\left({\mu }_n-1\right)+s_n& \mathrm{if}{\sigma }_n=1\\ 0& \mathrm{if}{\sigma }_n=0\end{array}$

Significance patterns are coded using a combination of two different techniques: an adaptive run-length code (MEL code) and a set of non-adaptive variable length coding (VLC) codes (CxtVLC codes).

Each sample has an exponent E_n which corresponds to the smallest number of least significant bits from v_n that are required to capture all of its non-zero bits. That is, E_n is the smallest non-negative integer E such that v_n<2^E.

An exponent upper bound U_q is coded for each quad q, via a predictor K_q and an “unsigned prediction residual” u_q≥0, where the predictor K_q for quad q is derived from previous coded sample magnitudes. Specifically, U_q≥E, where 25 is the largest exponent E n amongst all samples in the quad. We note that the value of ET' need not be explicitly computed by an encoder or decoded by a decoder, but it provides a convenient means to explain the present invention.

The significance pattern ρ_q and unsigned prediction residuals u_q for a quad are coded jointly, using a VLC coding scheme that involves two sub-codes, one of which (CxtVLC code) is dependent on a neighbourhood significance context C_q and best suited to table-lookup approaches, while the other (U-VLC code) is amenable to direct computation if required.

The CxtVLC also encodes a variable amount of additional information about the magnitude exponents of each sample in the quad. This information, known as EMB pattern information, is combined with the exponent bound U_q and significance pattern ρ_q to determine the number of sign and least significant magnitude bits from each sample of each quad that are packed into a MagSgn bit stream.

Specifically, the CxtVLC encodes the 4-bit quantity ρ_q, a 1-bit quantity u_q^off that is 1 if and only if u_q#0, and 0 to 4 bits of the EMB pattern ϵ_q, whose 4 bits correspond to the quantities e_n for each sample in the quad, where

$e_n=\lfloor \frac{{\nu }_n}{2^{U_q-1}}\rfloor$

The exponent bound U_q is an upper bound for the exponent E_n associated with each sample in quad q, so the e_n values produced by the above equation are certain to be binary valued—i.e., e_nϵ{1,1}.

The set of bits from the EMB pattern ϵ_q that are actually coded with the CxtVLC codeword for quad q is data dependent, being constrained in part by the maximum length of CxtVLC codewords, which is 7 in the HT Cleanup algorithm of ISO/IEC 15444-15:2019. Binary value k_n identify whether or not e_n is coded in the CxtVLC codeword. Decoders can deduce the k_n values, along with quantities i_n=k_n·e_n, from the CxtVLC codewords.

The CxtVLC mechanism can thus be understood as a variable-to-variable coding scheme, whereby a variable number of bits (e.g. 0 to 7 bits) are used to encode a variable amount of information about each quad (0 to 4 EMB bits, in addition to ρ_q and u_q^off). A separate set of codewords is used for each context c_q, which depends upon the significance of samples neighbouring the quad that have already been coded.

The least significant m_n=σ_n·(U_qk_n) bits of v_n are packed into the MagSgn bit-stream. Decoders can readily combine these bits with the values σ_n, k_n and i_n to recover v_n and hence the magnitude μ_n and sign s_n of each sample.

The HT Cleanup algorithm described in ISO/IEC 15444-15:2019 allows EMB pattern information to be coded via CxtVLC codewords only when u_q^off=1. This restriction was because it simplifies the encoding procedure, since when u_q^off=1 ,the encoder can always arrange for U_q to be equal to E_q^max, so that at least one of the e_n values is non-zero. This allows the encoder to avoid passing u_q^off explicitly to the lookup table used to implement CxtVLC encoding. Instead, it can look up CxtVLC codewords using only c_q, ρ_q and the 4-bit modified EMB pattern ϵ_q=u_q^off·ϵ_q—note that u_q^off is equal to 1 if and only if ϵ_q a is non-zero.

There is an ongoing need to provide methods which can improve coding efficiency in a manner compatible with standards compliant decoders. In particular, it is desirable to increase the amount of information that is jointly coded using redundancy-exploiting techniques such as variable length coding and it is desirable to do this without significantly altering the standardised decoding procedures themselves.

SUMMARY OF INVENTION

There can be benefits to allowing at least some EMB pattern information to be communicated via CxtVLC codes even when u_q^off=0. One important example where this can improve coding efficiency is when ρ_q=15 and U_q=E_q^max. In this case, all samples in a quad are significant and the EMB pattern ϵ_q can be interpreted as the most significant magnitude bit-plane associated with the quantities max{0,μ_n-1} for the quad's 4 samples. The predictor K_q is designed to be one less than the number of bits we would expect to require to represent v_n, so it is quite likely that all four samples in the quad have e_n=1 when ρ_q=15 and U_q=K_q; this can be efficiently captured via suitable CxtVLC codes.

This invention describes a mechanism to overcome the encoder complexity related limitations to coding EMB pattern information when u_q^off=0, allowing the development of CxtVLC codebooks with potentially higher coding efficiency.

An aspect provides method of image compression involving a variable-to-variable coding scheme, whereby a variable number of bits are used to encode a variable amount of information from a bounding magnitude bit-plane (EMB pattern) within a group of samples (hereinafter identified as a quad), in addition to encoding the significance of each sample in the quad (significance pattern) and a 1-bit unsigned residual offset value that indicates whether said bounding bit-plane is greater than a reference value (offset=1) or equal to the reference value (offset=0), the method involving a lookup table that the encoder addresses using a potentially modified version of the EMB pattern, a potentially modified version of the significance pattern, and any additional context information, in order to find the associated variable length codeword, wherein the modified significance pattern and modified EMB pattern, each consisting of 1 bit per sample in the quad, implicitly identify the quad's original significance pattern and the quad's 1-bit unsigned residual offset value, together with some or all of the quad's original EMB pattern, both when the unsigned residual offset value is 1 and, for specific values of the significance pattern, when the unsigned residual offset value is 0.

Another aspect provides a method of image compression providing an extension of a HT Cleanup encoding algorithm, as described in ISO/IEC 15444-15:2019, wherein context dependent variable length code (CxtVLC) lookup tables are indexed using a modified significance pattern, a modified EMB pattern and a context value, where the modified significance pattern and modified EMB pattern, each consisting of 4 bits, implicitly identify a quad's original 4-bit significance pattern and the quad's 1-bit unsigned residual offset value, together with some or all of the quad's original 4-bit EMB pattern, both when the unsigned residual offset value is 1 and, for specific values of the significance pattern, when the unsigned residual offset value is 0.

Another aspect provides a method of image wherein the context dependent variable length (CxtVLC) code includes codewords that communicate EMB pattern information for specific values of a quad's significance pattern when an unsigned residual offset value is 0.

In some embodiments the modified significance and modified EMB pattern values identify the original EMB pattern when the unsigned residual offset value is 0 and all samples in the quad are significant.

In some embodiments a quad's modified EMB pattern and modified significance pattern depend on whether or not the following three conditions hold:

• • a) all samples in the quad are significant;
  • b) the quad's unsigned residual offset value is 0; and
  • c) the quad's original EMB pattern is non-zero; such that the modified EMB pattern is set to 0 or its original EMB pattern, the latter being the case if the quad's unsigned residual offset value is 1 or else all three conditions hold, while the quad's modified significance pattern is set to its original significance pattern or 0, the latter being the case if all three conditions hold.

In some embodiments, the CxtVLC code includes multiple codewords for which the unsigned residual offset value is 0 and the significance pattern has all bits equal to 1, indicating that all samples in the quad are significant, at least one such codeword serving to communicate EMB pattern information for the quad.

In some embodiments a quad's modified EMB pattern is the same as its original EMB pattern, while the quad's modified significance pattern is set to either its original significance pattern or the bit-wise exclusive OR of its original significance and EMB patterns, the latter being the case if the quad's unsigned residual offset value is 0.

Embodiments also provide a method for decompression of images compressed in accordance with the embodiments above.

An aspect provides method of image decompression involving a variable-to-variable coding scheme, whereby a variable number of bits are used to encode a variable amount of information from a bounding magnitude bit-plane (EMB pattern) within a group of samples (hereinafter identified as a quad), in addition to encoding the significance of each sample in the quad (significance pattern) and a 1-bit unsigned residual offset value that indicates whether said bounding bit-plane is greater than a reference value (offset=1) or equal to the reference value (offset=0), the method involving a lookup table that the encoder addresses using a potentially modified version of the EMB pattern, a potentially modified version of the significance pattern, and any additional context information, in order to find the associated variable length codeword,

• • wherein the context dependent variable length (CxtVLC) code includes codewords that communicate EMB pattern information for specific values of a quad's significance pattern when an unsigned residual offset value is 0, wherein at least one such codeword serves to communicate magnitude exponent (EMB) pattern information for the quad.

Embodiments can provide an extension of the HT Cleanup encoding and decoding algorithm, as described in ISO/IEC 15444-15:2019.

DETAILS OF THE PREFERRED EMBODIMENT

Embodiments of the present invention relate to a modified coder implementation providing an extension for the High Throughput JPEG 2000 standard block coding algorithm described in the background section (the HTJ2K standard). Embodiments also provide an enhanced decoding method by virtue of the enhanced encoding.

In the method of embodiments of this invention, an enhanced block encoder addresses the CxtVLC table using context c_q, a modified significance pattern ρ_q, and a modified 4-bit pattern ϵ_q, which are determined in such a way that the value of u_q^off is implied by ρ_q and ϵ_q, while preserving all information in ρ_q and some or all of the information in ϵ_q, both when u_q^off=1 and also when u_q^off=0 and ρ_q=15.

Importantly, in this invention the encoder uses the same number of index bits to address the CxtVLC table as it did in the original HT block coding algorithm—4 bits from ρ_q, 4 bits from ϵ_q, plus the bits associated with the context c_q. In this way, the enhancement of this invention can be obtained with almost no increase in implementation complexity.

The decoding procedure is not actually changed at all by the method of this invention. CxtVLC decoding still produces values for ρ_q, u_q^off and the k_n and i_n values, from which the decoder knows the number of bits m_n=σ_n·(U_q-k_n) to unpack from the MagSgn bit-stream, combining them with σ_n and i_n to recover v_n and thence μ_n and s_n. However, the CxtVLC table itself will be different, since it can now include codewords corresponding to u_q^off=0, in which one or more of the k_n are non-zero.

In a preferred embodiment, the encoder constructs p q and E q as follows:

$\stackrel{=}{{ϵ}_q}=\{\begin{array}{cc}{ϵ}_q& \mathrm{if}{\rho }_q=15\mathrm{and}{u}_q^{\mathrm{off}}=0\mathrm{and}{ϵ}_q\ne 0\\ u_q^{\mathrm{off}}·{ϵ}_q& \mathrm{otherwise}\end{array}$ $\stackrel{=}{{\rho }_q}=\{\begin{array}{cc}0& \mathrm{if}{\rho }_q=15\mathrm{and}{u}_q^{\mathrm{off}}=0\mathrm{and}{ϵ}_q\ne 0\\ {\rho }_q& \mathrm{otherwise}\end{array}$

This works because in the case where ϵ_q#0, the significance pattern ρ_q cannot possibly be 0. It can be deduced, therefore, that when ϵ_q≠0 and ρ_q=0 the value of ρ_q must be 15. Moreover, u_q^off=1 if and only if ϵ_q≠0 and ρ_q≠0, so long as the encoder makes the sensible choice U_q=E_q^max whenever K_q≤E_q^max so that ϵ_q≠0 whenever u_q^off=1. In this way, non-zero EMB patterns ϵ_q can be made available for CxtVLC coding even when u_q^off=0, at least in the important case where ρ_q=15.

More generally, embodiments of the invention can exploit the fact that whenever any EMB bit e_n=1, the corresponding significance bit must be a σ_n=1, so that the corresponding bit within ρ_q a can be used to implicitly communicate u_q^off.

In one general embodiment of the invention, that exploits this principle, the encoder constructs ρ_q and ϵ_q as follows:

$\stackrel{=}{ϵ}={ϵ}_q$ ${\stackrel{=}{\rho }}_q=\{\begin{array}{cc}{\rho }_q\oplus {ϵ}_q& \mathrm{if}{u}_q^{\mathrm{off}}=0\\ {\rho }_q& \mathrm{otherwise}\end{array}$

Here ⊕ denotes bit-wise “exclusive OR.” In this embodiment, u_q^off is implicitly identified as 0 if ϵ_q==0 or if any bit position within ϵ_q is 1, while the corresponding bit position within ρ_q is 0; otherwise u_q^off must be 1. As before, this relies on the assumption that the encoder makes the sensible choice U_q=E_q^max whenever K_q≤_q^max so that ϵ_q≠0 whenever u_q^off=1. Also, ρ_q is easily recovered from the modified EMB and significance bit patterns by taking their bit-wise “inclusive OR” —i.e., ρ_q=ρ_qVϵ_q.

To summarise key points described above:

• • The CxtVLC coding procedure in the HTJ2K standard allows a decoder to decode 3 quantities (significance pattern ρ_q, partial EMB pattern information, and unsigned residual offset u_q^off), all from a variable length codeword with 0-7 bits.
  • Importantly, the amount of EMB pattern information contained in a CxtVLC codeword is variable—anywhere from no EMB information to all 4 bits of EMB pattern information are communicated, depending on the codeword—it is a variable-to-variable coding scheme, coding a variable amount of information, using a variable number of bits.
  • The HTJ2K encoding procedure only uses a 3-bit context value, a 4-bit significance pattern ρ_q and a 4-bit EMB pattern ϵ_q, multiplied by the binary u_q^off value. These 11 bits address a lookup table that generates the CxtVLC codewords, relying upon the fact that these implicitly identify u_q^off=0, but at the expense of discarding all EMB pattern information ϵ_q when u_q^off=0.
  • The present invention is based on the insight that there are ways to communicate some EMB pattern information even when u_q^off=0 without using additional bits to address the CxtVLC table in the encoder.

The key to this is to address the lookup table using a modified EMB pattern and a modified significance pattern, having the same number of bits as the original EMB and significance patterns.

The terminology of relevance to the description above is summarised as follows:

• • 1. EMB pattern—the bit pattern ϵ_q consisting of all e n bits for a given quad q.

“EMB pattern” is a term in the HTJ2K standard, where it is a 4-bit word for each 22 quad being coded, that consists of the bits e_n. This is the most significant bit-position that can contain non-zero bits. The EMB pattern is statistically skewed and in the context of the present invention presents an opportunity for coding efficiency gains through variable length coding methods, which is why it is desirable to include EMB pattern information in the variable length CxtVLC codewords whenever possible, subject to the constraint that CxtVLC codewords should exceed a given maximum length.

• • 2. Modified EMB pattern—a bit pattern ϵ_q, derived from and having the same number of bits as the original EMB pattern ϵ_q, that is used to address the encoder's CxtVLC lookup table.
  • 3. Partial EMB pattern information—partial information about the original EMB pattern (0 or more of the original EMB bits) that the decoder discovers when it encounters a particular CxtVLC codeword.
  • 4. Significance pattern—the bit pattern ρ_q consisting of all σ_n bits for a given quad q, identifying whether or not each sample in the quad is non-zero.
  • 5. Modified significance pattern—a bit pattern ρ_q, derived from and having the same number of bits as the original significance pattern, that is used to address the encoder's CxtVLC lookup table.

It will be understood to persons skilled in the art of the invention that many modifications may be made without departing from the spirit and scope of the invention.

In the claims which follow and in the preceding description of the invention, except where the context requires otherwise due to express language or necessary implication, the word “comprise” or variations such as “comprises” or “comprising” is used in an inclusive sense, i.e. to specify the presence of the stated features but not to preclude the presence or addition of further features in various embodiments of the invention.

It is to be understood that, if any prior art publication is referred to herein, such reference does not constitute an admission that the publication forms a part of the common general knowledge in the art, in Australia or any other country.

Claims

1. A method of image compression involving a variable-to-variable coding scheme, whereby a variable number of bits are used to encode a variable amount of information from a bounding magnitude bit-plane (EMB pattern) within a group of samples (hereinafter identified as a quad), in addition to encoding the significance of each sample in the quad (significance pattern) and a 1-bit unsigned residual offset value that indicates whether said bounding bit-plane is greater than a reference value (offset=1) or equal to the reference value (offset=0), the method involving a lookup table that the encoder addresses using a potentially modified version of the EMB pattern, a potentially modified version of the significance pattern, and any additional context information, in order to find the associated variable length codeword, wherein the modified significance pattern and modified EMB pattern, each consisting of 1 bit per sample in the quad, implicitly identify the quad's original significance pattern and the quad's 1-bit unsigned residual offset value, together with some or all of the quad's original EMB pattern, both when the unsigned residual offset value is 1 and, for specific values of the significance pattern, when the unsigned residual offset value is 0.

2. The method of image compression of claim 1, wherein the context dependent variable length (CxtVLC) code includes codewords that communicate EMB pattern information for specific values of a quad's significance pattern when an unsigned residual offset value is 0.

3. The method of image compression of claim 1 wherein the modified significance and modified EMB pattern values identify the original EMB pattern when the unsigned residual offset value is 0 and all samples in the quad are significant.

4. The method of image compression of claim 2, where a quad's modified EMB pattern and modified significance pattern depend on whether or not the following three conditions hold: a) all samples in the quad are significant;b) the quad's unsigned residual offset value is 0; andc) the quad's original EMB pattern is non-zero;such that the modified EMB pattern is set to 0 or its original EMB pattern, the latter being the case if the quad's unsigned residual offset value is 1 or else all three conditions hold, while the quad's modified significance pattern is set to its original significance pattern or 0, the latter being the case if all three conditions hold.

5. The method of image compression of claim 1, where the CxtVLC code includes multiple codewords for which the unsigned residual offset value is 0 and the significance pattern bits are all 1, indicating that all samples in the quad are significant, at least one such codeword serving to communicate EMB pattern information for the quad.

6. The method of image compression of claim 1 wherein, a quad's modified EMB pattern is the same as its original EMB pattern, while the quad's modified significance pattern is set to either its original significance pattern or the bit-wise exclusive OR of its original significance and EMB patterns, the latter being the case if the quad's unsigned residual offset value is 0.

7. The method of image compression as claimed in claim 1, wherein the method provides an extension of a HT Cleanup encoding algorithm, as described in ISO/IEC 15444-15:2019.

8. A method of image decompression involving a variable-to-variable coding scheme, whereby a variable number of bits are used to encode a variable amount of information from a bounding magnitude bit-plane (EMB pattern) within a group of samples (hereinafter identified as a quad), in addition to encoding the significance of each sample in the quad (significance pattern) and a 1-bit unsigned residual offset value that indicates whether said bounding bit-plane is greater than a reference value (offset=1) or equal to the reference value (offset=0), the method involving a lookup table that the encoder addresses using a potentially modified version of the EMB pattern, a potentially modified version of the significance pattern, and any additional context information, in order to find the associated variable length codeword, wherein the context dependent variable length (CxtVLC) code includes codewords that communicate EMB pattern information for specific values of a quad's significance pattern when an unsigned residual offset value is 0, wherein at least one such codeword serves to communicate magnitude exponent (EMB) pattern information for the quad.

9. The method of image decompression as claimed in claim 8, wherein the method provides an extension of a HT Cleanup encoding and decoding algorithm, as described in ISO/IEC 15444-15:2019.