CN102223534B - All-parallel bit plane coding method for image compression - Google Patents

Abstract

The present invention relates to a fully parallel bit-plane encoding method for image compression, in particular to a high-speed full-parallel encoding of bit-planes and encoding channels that complies with the Joint Photographic Experts Group (JPEG) 2000 image compression standard, low memory consumption Bit-plane coding method; this invention uses a single column of bits as a data unit, and only needs to cache the significance status information of the previous column of bits, and read the original data of the current column and the following two columns, and the current The channel of the column and the bit plane are encoded in parallel; after that, only a new column of data needs to be read each time to realize the encoding cycle; the implementation structure of the encoding method only needs 259 cycles to process 32

32 wavelet subbands.

Description

Fully Parallel Bit-Plane Coding Method for Image Compression

technical field

The invention relates to a fully parallel bit-plane coding method for image compression, in particular to a high-speed full-parallel coding of bit-planes and coding channels that conforms to the Joint Photographic Experts Group (JPEG) 2000 image compression standard, low memory consumption Bit-plane encoding method.

Background technique

At present, the latest international image compression standard JPEG2000 (reference [1]) has the qualities of intra-frame coding, good compression effect, support for region-of-interest coding and progressive transmission, and thus has been promoted in high-end digital image applications. In the process of image compression, the original image of two-dimensional discrete wavelet transform (DWT) with lifting structure is firstly used, and then the bit plane coder (BPC) is used to process the subbands of wavelet coefficients to generate context and code values (CX/D), and then sent to arithmetic coding device (AC) for compression coding. Since BPC needs to establish a context for each bit of the discrete wavelet transform result of the image, the computational complexity is extremely high, and a large amount of memory is consumed. With the popularity of various high-definition digital images, how to improve the full parallelism of the bit-plane coding method and effectively reduce memory consumption has become one of the research hotspots in this field.

The BPC method used in the JPEG2000 image compression standard was proposed by David Taubman. This method organizes the quantized P-bit signed wavelet coefficients into N×N code blocks, and then divides the signs of the wavelet coefficients into one An independent sign plane, which divides the bit order of the wavelet coefficient into (P-1) layers of N×N bit planes; in a single bit plane, every 4 rows of bits are divided into a strip. The encoder starts from the highest bit plane, top-to-bottom, left to right, and encodes bit-by-bit, bit-by-bit, bit-by-bit, according to the order of the raster.

The bit encoding algorithm in a single band is as follows:

1) Judging whether the current bit belongs to the significance propagation channel (SPP): if the bit is currently insignificant, but at least one of the 8 adjacent bits is significant, then it belongs to the channel and zero-coded (ZC ) and calculate the context. Then judge whether it is significant; if so, update the significance state and perform symbol coding (SC); otherwise, do not. If none of the adjacent areas of the bit is significant or the bit itself is already significant, then it does not belong to the channel, and this point is skipped to judge the next point. Perform this operation on all bits in the strip, and then return to the starting point to perform 2);

2) Judging whether the current bit belongs to the amplitude refinement channel (MRP): if the bit is currently significant and unencoded, it belongs to the channel, and the amplitude refinement encoding is performed; otherwise, it does not belong to the channel, skip this point to judge next point. Perform this operation on all bits in the strip, and then return to the starting point to perform 3);

3) Execution of a clear pass (CLP): use run-length coding (RLC), zero coding, and sign coding to encode all remaining bits that do not belong to the saliency propagation pass and the magnitude refinement pass.

The standard BPC algorithm starts from the highest bit plane, from top to bottom, from left to right, bit-by-bit plane-by-strip, bit-by-bit, and codes channel-by-channel in the order of SPP channel, MRP channel, and CLP channel. This sequential encoding method is very slow, and storage requires a large amount of intermediate data. Therefore, how to effectively improve the encoding speed of bit-plane encoders and reduce internal storage consumption has become one of the academic hotspots concerned by scholars at home and abroad.

The representative research results in this field are as follows: Document [1] firstly proposed a bit-plane coding algorithm based on pixels, which was adopted by the JPEG2000 standard; Document [2] adopted a state machine structure to realize bit-plane coding; Document [3] ] proposed a column-based encoding channel parallel structure, that is, two window encoders were used to simultaneously scan and encode three channels. Due to the sequential order of channel scanning, the second window can start encoding only after the saliency state of the first window is updated, and the two window encoders need to interact, which increases the control complexity; literature [4] proposed Bit skipping and column skipping algorithms, that is, skipping bits and columns of output without context to reduce the number of calculations. This method cannot take advantage of circuit parallelization, the degree of optimization depends on the image data, and the performance fluctuates greatly. Literature [5] proposed a bit-plane parallel algorithm, which realized parallelism between encoded bit-planes. However, the encoding unit of a single bit plane adopts the schemes of [3] and [4] at the same time, which has high control complexity and consumes a lot of resources. This shortcoming is particularly prominent after hardware multiplexing for parallel encoding of bit planes.

[1] Taubman D. High performance scalable image compression with EBCOT.IEEETrans.on Image Processing, 2000, 9(7): 1158-1170

[2] Andra K., Chakrabarti C., Acharya T., A high-performance JPEG 2000 architecture. IEEE Trans. on Circuits and System for Video Technology, 2003, 13(3): 209-218

[3] Chiang J.-S., Lin Y.-S., Hsieh C.-Y., Efficient pass-parallel architecture for EBCOTin JPEG 2000.In: IEEE International Symposium on Circuits and Systems, 2002.773-776

[4] Lian C.-J., Chen K.-F., Chen H.-H., et al. Analysis and architecture design of block-coding engine for EBCOT in JPEG 2000. IEEE Trans. on Circuits and System for Video Technology , 2003, 13(3): 219-230

[5] Liu K., Li Y.-S., Wu C.-K., A high performance EBCOT coding and its VLSIarchitecture. Journal of Software, 2006, 17(7): 1553-1560

Contents of the invention

Technical problem: The present invention relates to a fully parallel bit-plane coding method for image compression, by removing the correlation between bit-planes, parallel and independent coding of each bit-plane is realized; in a single bit-plane, a single row of bits is used as The data unit only caches the significance status information of the previous column of bits, and reads the original data of the current column and the next two columns, and completes the parallel encoding of the channel and bit plane of the current column within one encoding window; after that, it only needs to read each time Enter a new column of data to realize the encoding cycle.

The described removal of the correlation between the bit planes to realize the parallel independent encoding of each bit plane means that after reading in a column of wavelet coefficients, the wavelet coefficients read in are converted based on formulas 1 and 2 to generate The significance state s of all bit planes and the first amplitude refinement state mrf, thereby removing the correlation between the bit planes; each bit plane can be encoded independently and in parallel by using the generated s and mrf;

${\mathrm{the\; s}}_l^{\mathrm{no}}=\underset{i=l+1}{\overset{\mathrm{MSB}}{\mathrm{\Σ}}}{\mathrm{mag}}_i^{\mathrm{no}}$ Formula 1

表示第l位平面的第n列比特位的显著性状态，

表示第l位平面的第n列比特位的首次幅值细化状态，

表示第i位平面第n列比特位的幅度，MSB表示最高位平面的层数，∑表示或运算。In Equation 1 and Equation 2,

Indicates the significance state of the bits in the nth column of the lth bit plane,

Represents the first amplitude refinement state of the nth column bit of the lth bit plane,

Indicates the magnitude of the bit in the nth column of the i-th bit plane, MSB indicates the number of layers of the highest bit plane, and ∑ indicates an OR operation.

In a single bit plane, a single column of bits is used as a data unit, only the significance state information of the previous column of bits is cached, and the original data of the current column and the following two columns are read, and the current column is completed within one coding window The channels and bit planes are encoded in parallel; after that, only one column of new data needs to be read each time to realize the encoding cycle, which means:

其中

表示第(n-1)列比特位经显著性传播通道SPP编码更新后的显著性状态，

表示第n列比特位原始显著性状态；对第n列比特位进行幅值细化通道MRP判决时，其本身及前后列显著性状态分别为

对当前列进行清除通道CLP判决时，其本身及前后列显著性状态分别为

若比特位属于SPP或CLP，则可能采用零编码和符号编码进行编码，从而更新比特位的显著性状态；若比特位属于MRP，则必然采用幅值细化编码进行编码，不会对比特位的显著性状态进行更新；推得公式3：The standard bit-plane coding BPC algorithm stipulates that when the SPP decision is made on the nth column of bits in a single bit-plane, the significance states of itself and adjacent columns are respectively

in

Indicates the saliency state of the (n-1)th column bit after being updated by the saliency propagation channel SPP code,

Indicates the original saliency state of the bit in the nth column; when the MRP judgment of the amplitude refinement channel is performed on the bit in the nth column, the saliency state of itself and the preceding and following columns are respectively

When making a clear channel CLP judgment on the current column, the saliency status of itself and the front and rear columns are respectively

If the bit belongs to SPP or CLP, it may use zero coding and sign coding to update the significance state of the bit; if the bit belongs to MRP, it must use amplitude refinement coding to code, and will not Update the salience state of ; deduce the formula 3:

${\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{mrp}}^{\mathrm{no}}={\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{spp}}^{\mathrm{no}}$ Formula 3

According to formula 3, the condition for updating the saliency state of the bits in the nth column before and after the judgment of the three channels can be summarized as:

$\left\{\begin{array}{c}{\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{spp}}^{\mathrm{no}}=f({\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{spp}}^{\mathrm{no}-1},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}+1})\\ {\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{mrp}}^{\mathrm{no}}=f({\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{spp}}^{\mathrm{no}-1},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}+1},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}+2})\\ {\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{clp}}^{\mathrm{no}}=f({\mathrm{the\; s}}_{\mathrm{update}\_\mathrm{clp}}^{\mathrm{no}-1},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}+1},{\mathrm{the\; s}}_{\mathrm{original}}^{\mathrm{no}+2})\end{array}\right.\mathrm{no}\&Greater\; Equal;1$ Formula 4

用于预测第(n+1)列在SPP编码后的显著性状态 In formula 4,

Used to predict the significance status of the (n+1)th column after SPP encoding

The standard bit-plane coding algorithm stipulates that the saliency state outside the left and right borders of the strip is 0 by default, then the saliency state update conditions of the bits in column 0 before and after three-channel coding are:

$\left\{\begin{array}{c}{\mathrm{the\; s}}_{\mathrm{<figure-callout\; id="0"\; label="update\; \_\; spp"\; filenames="110818112806.png,110818112809.png"\; state="\{\{state\}\}">update</figure-callout>}\_\mathrm{spp}}^0=f(0,{\mathrm{the\; s}}_{\mathrm{original}}^0,{\mathrm{the\; s}}_{\mathrm{original}}^1)\\ {\mathrm{the\; s}}_{\mathrm{<figure-callout\; id="0"\; label="update\; \_\; mrp"\; filenames="110818112806.png,110818112809.png"\; state="\{\{state\}\}">update</figure-callout>}\_\mathrm{mrp}}^0=f(0,{\mathrm{the\; s}}_{\mathrm{original}}^0,{\mathrm{the\; s}}_{\mathrm{original}}^1,{\mathrm{the\; s}}_{\mathrm{original}}^2)\\ {\mathrm{the\; s}}_{\mathrm{<figure-callout\; id="0"\; label="update\; \_\; clp"\; filenames="110818112806.png,110818112809.png"\; state="\{\{state\}\}">update</figure-callout>}\_\mathrm{clp}}^0=f(0,{\mathrm{the\; s}}_{\mathrm{original}}^0,{\mathrm{the\; s}}_{\mathrm{original}}^1,{\mathrm{the\; s}}_{\mathrm{original}}^2)\end{array}\right.\mathrm{no}=0$ Formula 5

构成一个编码窗口；若同时已知第(n-1)列的

和

作为辅助列，即可根据公式3和公式4完成第n列的编码通道并行编码，并将

和

缓存作为新的辅助列；之后，编码窗口只需读入一列新的小波系数即可实现编码算法的循环；而编码窗口初始编码时所需的首列显著性状态辅助信息则可以通过公式5求得。After reading the wavelet coefficients of column n, column (n+1) and column (n+2), the values of each bit plane are obtained by formula 1 and formula 2

Constitute a coding window; if the (n-1)th column is known at the same time

and

As an auxiliary column, the parallel encoding of the encoding channel of the nth column can be completed according to formula 3 and formula 4, and

and

cache as a new auxiliary column; after that, the encoding window only needs to read in a new column of wavelet coefficients to realize the cycle of the encoding algorithm; and the first column of saliency state auxiliary information required for the initial encoding of the encoding window can be obtained by formula 5 have to.

Beneficial effects: this algorithm realizes bit plane and channel parallelism at the same time, and its advantage lies in that it is easy to implement corresponding hardware pipeline operation; only one coding window is needed, the calculation and control are simple, the storage capacity is small, and small resources can be kept in hardware implementation Overhead; can effectively reduce the complexity of input and output operations between DWT and BPC. After the BPC takes out the three columns of wavelet coefficients, it can simultaneously complete the CX/D calculation of the first column in all bit planes, and cache the saliency status of the current column after SPP encoding and CLP encoding. Afterwards, a new column of wavelet coefficients is read in each cycle, and the bit-plane encoding of the next column can be performed. Corresponding to the pipeline operation in the hardware implementation, it is very easy to realize. In a single bit plane, only the auxiliary parameters required by the current encoding window need to be stored, and no more buffer consumption is required. The encoding window only encodes one column of data at a time, with a small amount of calculation, and the data reading relies on pipeline operations, so the control complexity is low. Therefore, resource consumption is kept low when implemented in hardware. This quality is especially important because of the need to multiplex the hardware units of a single bit-plane encoder when implementing bit-plane parallel encoding. Using this algorithm, after reading in three columns of wavelet coefficients, BPC can immediately operate each column of data taken out from DWT to complete encoding and output without redundant storage and control, which greatly improves the data input and output between BPC and DWT s efficiency.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

In the BPC algorithm in Fig. 1, a schematic diagram of scanning and coding of strips and bits in raster order.

Fig. 2 is a schematic diagram of the principle of implementing three encoding channels in parallel through one encoding window.

Fig. 3 Schematic diagram of bit-plane parallel encoding.

Fig. 4 is a schematic diagram of the implementation structure of the fully parallel bit-plane encoding method.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and the implementation examples given by the inventor.

The present invention first removes the correlation between each bit plane, that is, obtains the significance state s of each bit plane and the first amplitude refinement state mrf through the state parameter generation unit; after being buffered by the register buffer unit, in a single bit plane, the buffer By reading the significance status information of the previous column of bits and reading the original data of the current column and the next two columns, the parallel encoding of the channel and bit plane of the current column can be completed within one encoding window; after that, only one new column needs to be read each time The data can realize the encoding cycle.

The specific implementation examples are as follows:

1) Read in a list of wavelet coefficients, convert the read-in wavelet coefficients based on formulas 1 and 2, generate the salience state and the first amplitude refinement state of all bit planes except the sign bit plane and the highest bit plane, and send to Each bit plane is encoded in parallel; the encoding structure and encoding process of each bit plane are the same;

2) The encoding window output buffers the programmed (n-1)th column saliency state parameters, the first amplitude refinement state parameters, amplitude and sign as auxiliary, and reads in the new (n+2) saliency state Parameters, the first amplitude refinement state parameters, amplitude and sign, plus the existing nth column and (n+1)th column in the encoding window, start encoding the nth column;

3) Judging the channel to which each bit belongs according to the significance state parameter, if the significance of a certain bit is "0", and the significance of any one of the bits adjacent to it is "1", then this bit belongs to SPP , if the significance of a certain bit is "1", then this bit belongs to MRP, and other bits belong to CLP. After the judgment is completed, each bit is encoded according to different channels, and the encoding of each channel is parallel;

4) For the SPP channel, use ZC to calculate CX/D for bits with an amplitude of "0", use ZC and SC to calculate CX/D for bits with an amplitude of "1", and update each bit of the SPP channel Significance and register for use in SPP channel for the next column of data; for MRP channel, first predict the significance of each bit after passing through the SPP channel, and update the significance of each bit, and then use MRC to calculate each bit For the CX/D of the bit, the significance of each bit of the MRP channel is updated for use in the MRP channel for the next column of data; for the CLP channel, the significance of each bit after the SPP channel is first predicted, and each bit The significance of the bit is updated, and then when the significance of itself is "0" and the significance of its adjacent bits is all "0", use RLC to calculate CX/D. For other bits, if the bit If the bit amplitude is "0", use ZC to calculate CX/D, if the amplitude is "1", use ZC and SC to calculate CX/D, and update the significance of each bit of the CLP channel for CLP on the next column of data channel used.

5) After encoding the nth column, return to 1), and start to read a new column to continue cyclic encoding.

Claims (2)

1. a full parallel bit plane coding method that is used for image compression is characterized in that removing the correlation of using the bit plane that obtains in the JPEG2000 image compression process, realizes the parallel absolute coding of each bit plane; In single bit plane, as data cell, the conspicuousness state information of buffer memory previous column bit only, and read initial data when prostatitis and follow-up two row is finished passage and bit plane parallel encoding when the prostatitis in a coding window with single-row bit; Only need read in afterwards the new data of row at every turn, realize the coding circulation;

The described correlation of having removed between bit plane, realize the parallel absolute coding of each bit plane, after referring to read in a row wavelet coefficient, change based on formula 1 and 2 pairs of wavelet coefficients that read in, generate conspicuousness state s and the first amplitude refinement state mrf on divided-by symbol bit plane and out-of-plane all the bit planes of highest order, thereby removed the correlation between bit plane; Each bit plane uses s and the mrf that generates, and gets final product the coding of independent parallel;

$s_l^n=\underset{i=l+1}{\overset{\mathrm{MSB}}{\mathrm{\&Sigma;}}}{\mathrm{mag}}_i^n$ Formula 1

Formula 2

In formula 1 and the formula 2,

The conspicuousness state that represents the n row bit of l bit plane,

The first amplitude refinement state that represents the n row bit of l bit plane,

Represent the amplitude of i bit plane n row bit, MSB represents the number of plies on highest order plane, and Σ represents exclusive disjunction.

2. a kind of full parallel bit plane coding method for image compression as claimed in claim 1, it is characterized in that described in single bit plane, with single-row bit as data cell, the conspicuousness state information of buffer memory previous column bit only, and read when prostatitis and follow-up two initial data that are listed as, in a coding window, finish passage and bit plane parallel encoding when the prostatitis; Only need read in afterwards the new data of row at every turn, realize the coding circulation, refer to:

Standard bit plane coding BPC algorithm dictates, when n row bit in the bit-plane was carried out the SPP judgement, the conspicuousness state of itself and front and back adjacent column was respectively

, ,

Wherein

The conspicuousness state of expression (n-1) row bit after conspicuousness propagation ducts SPP coding upgrades,

Represent the original conspicuousness state of n row bit; When n row bit was carried out amplitude refinement MRP judgement, itself and front and back row conspicuousness state were respectively

,

,

When adjudicating remove channel C LP when the prostatitis, itself and front and back row conspicuousness state are respectively

,

,

If bit belongs to SPP or CLP, then may adopt Zero-code and symbolic coding to encode, thus the conspicuousness state of update bit position; If bit belongs to MRP, then must adopt the amplitude refinement to encode, can the conspicuousness state of bit not upgraded; Push away to get formula 3:

$s_{\mathrm{update}\_\mathrm{mrp}}^n=s_{\mathrm{update}\_\mathrm{spp}}^n$ Formula 3

Can be summarized as according to formula 3, the n row bits conspicuousness state update condition before and after three passage judgements:

$\left\{\begin{array}{c}{s}_{\mathrm{update}\_\mathrm{spp}}^n=f(s_{\mathrm{update}\_\mathrm{spp}}^{n-1},s_{\mathrm{original}}^n,s_{\mathrm{original}}^{n+1})\\ s_{\mathrm{update}\_\mathrm{mrp}}^n=f(s_{\mathrm{update}\_\mathrm{spp}}^{n-1},s_{\mathrm{original}}^n,s_{\mathrm{original}}^{n+1},s_{\mathrm{original}}^{n+2})\\ s_{\mathrm{update}\_\mathrm{clp}}^n=f(s_{\mathrm{update}\_\mathrm{clp}}^{n-1},s_{\mathrm{original}}^n,s_{\mathrm{original}}^{n+1},s_{\mathrm{original}}^{n+2})\end{array}\right.n\&GreaterEqual;1$ Formula 4

In the formula 4,

Be used for the conspicuousness state after prediction (n+1) is listed in the SPP coding

Standard bit plane encryption algorithm regulation, the conspicuousness state beyond the border, the band left and right sides is defaulted as 0, and then the conspicuousness state update condition of the 0th row bit before and after three channel codings is:

$\left\{\begin{array}{c}{s}_{\mathrm{update}\_\mathrm{spp}}^0=f(0,s_{\mathrm{original}}^n,s_{\mathrm{original}}^1)\\ s_{\mathrm{update}\_\mathrm{mrp}}^0=f(0,s_{\mathrm{original}}^0,s_{\mathrm{original}}^1,s_{\mathrm{original}}^2)\\ s_{\mathrm{update}\_\mathrm{clp}}^0=f(0,s_{\mathrm{original}}^0,s_{\mathrm{original}}^1,s_{\mathrm{original}}^2)\end{array}\right.n=0$ Formula 5

Reading n row, (n+1) row behind the wavelet coefficient of (n+2) row, obtain each bit planes by formula 1 and formula 2 ,

,

, consist of a coding window; If known (n-1) is listed as simultaneously

With As supplementary column, can finish the coding pass parallel encoding of n row according to formula 3 and formula 4, and will

With

Buffer memory is as new supplementary column; Afterwards, coding window only need read in the circulation that the new wavelet coefficient of row can be realized encryption algorithm; First required during coding window initial code conspicuousness state supplementary then can be tried to achieve by formula 5.

Images