JP3281423B2 - Code amount control device at the time of image encoding - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION The present invention relates to an image encoding system in which 1 / S (sub-frame) of one frame, which is a constituent unit of an image, is used as a processing unit. Is related to a code amount control device at the time of image encoding in which fixed control is performed so that a predetermined value is set as an upper limit.
The present invention relates to a code amount control device at the time of image encoding that can be applied to moving image encoding for a TR or the like and still image encoding for a still image storage medium (for example, a digital still camera).

SUMMARY OF THE INVENTION The present invention relates to a moving picture coding for a moving picture storage medium (for example, a cassette type digital VTR) and a still picture for a still picture storage medium (for example, a digital still camera). In encoding, 1 / S of one frame, which is a constituent unit of an image,
(Hereinafter referred to as “sub-frame”. S is a natural number and S ≧ 1,
For example, one field in a TV signal is one field when S = 2.
This is a subframe. ) Is used as a processing unit, and when image encoding is performed such that one subframe is divided into a plurality of pixel blocks, a predetermined value is set as an upper limit to a code amount generated for each subframe which is the processing unit. In this manner, a fixed control method (fixed-length code control method) is used, and an optimum coding parameter _Qk is selected and encoded for each block, and the block is optimized to optimize the total code amount and the sum of quantization errors.

[0003]

[Prior art]

<< Necessity of Fixed-Length Coding in Moving Image Coding of Moving Image Storage Media >> Generally, in the transmission of moving images and recording of moving images without editing, individual frames / fields which are constituent units of images are used. Is not necessarily required to be constant for each frame / field.
Therefore, a moving picture coding method (MPE
G), the code amount is not always constant for each image, for example, by utilizing image correlation between frames / fields. However, a desired constant amount is set for the entire moving image to be encoded. A method of achieving a rate compression rate (constant rate code control method) is used.

In this method, a desired code rate is realized by writing the generated codes into a FIFO type buffer and reading them out at a desired rate constantly. In this case, the encoding parameters are controlled so that the FIFO type buffer does not cause underflow or overflow (ideally, the buffer occupancy is almost always 常 に). This control can be easily and reliably performed if the maximum capacity of the FIFO buffer is increased to some extent.

However, in a moving image storage medium such as a cassette type digital VTR, not only recording / reproducing of the entire moving image but also cutting out of a specific frame from the moving image and editing in a frame / field unit are required. Required. In such editing, it goes without saying that editing the encoded data recorded on the medium as it is without decoding it into an image can prevent image quality deterioration and shorten the editing execution time. Therefore, it is desirable to perform fixed-length encoding using a frame / field as a processing unit.

Further, in a digital VTR, one frame / field, which is a constituent unit of an image, is often recorded over a plurality of tracks, and an image corresponding to one track is used as a processing unit for image coding. It is also possible.

As is apparent from the above description, in moving picture coding for moving picture storage media, 1 / S (subframe, where S is a natural number and S ≧ 1) of one frame, which is a unit of picture, is used. It is an essential condition to adopt a system (fixed-length code control system) in which the encoding procedure is completed and the amount of code generated for each processing unit is fixed.

<< Conventional Fixed-Length Coding Technique >> Hereinafter, an image coding method is shown below in units of subframes. Orthogonal transform (DCT, etc.) + Scalar quantization + variable-length code (entropy code) (1) ) Will be described as an example.

In this case, the encoding is performed in units of m × n pixel blocks of m pixels vertically × n pixels horizontally in the subframe, and a variable length code is generated.

That is, m × n transform coefficients obtained by orthogonal transform are obtained for each block, and these are quantized to obtain m
A set of × n integers is obtained, and the set of integers is converted to a variable length code.

As can be seen from this procedure, how to quantize the transform coefficients is an important factor that determines both the image quality (quantization noise) and the amount of generated variable-length code (compression ratio). Become.

It is known that in orthogonal transform such as DCT for general images, energy is concentrated on low-order transform coefficients. Based on this fact, it is common to use a quantizer for performing the above-described quantization on the transform coefficients in such a manner that the lower-order transform coefficients are finely quantized and the higher-order transform coefficients are coarsely quantized.

However, since the degree of concentration of the energy on the low-order coefficients differs depending on the image to be encoded, if the quantizer is fixed, a code (variable-length code) finally obtained by the image is obtained. ) Are different. That is, in the case of a fine pattern, the code amount increases, and in the case of a pattern having many flat portions, the code amount decreases.

For this reason, in moving image coding for transmission and the like, in order to realize a constant code rate, as described above, a FIFO type buffer for storing codes is used.
A method of monitoring the buffer occupancy and controlling the "roughness" of the scalar quantizer is used. In this case, as an average value of the code amount over a plurality of frames / fields,
It is sufficient if a predetermined code rate can be realized.

[0015] However, in the fixed-length intra-subframe coding method in question here, the code amount per coding processing unit (1 subframe) must be strictly less than a predetermined code amount. Such buffer control is not available.

As described above, in the conventional method often used in fixed control, a "prediction function" for predicting the code amount is prepared from the result of orthogonal transformation for a pixel block, and based on that, the block is subjected to the block. This determines the "roughness" of the quantizer to be applied.

<< Example of Conventional Code Amount Prediction Function (Code Amount Prediction by Activity) >> A code amount prediction function often used in the past predicts the amount of entropy code per block of m × n pixels. In general, the activity A per block (the sum of squares of transform coefficients other than the DC component among the orthogonal transform coefficients, that is, AC energy) and a scalar serving as an encoding parameter Q for controlling the "roughness" of the quantizer. It is often expressed as a function of the quantity Q (eg, the Q factor in JPEG), f (Q, A) (2). This code amount prediction function f
Since there is no way to theoretically find (Q, A),
As a practical method, encoding is performed on a plurality of images considered to be standard with respect to several scalar amounts Q, and the activity A in each block and the amount b of the entropy code are calculated. By measuring and interpolating from a set of a large number of these measured values, {(A, Q, b)} (3), the code amount prediction function: b = f (A, Q) (4) Alternatively, this function is solved for a scalar quantity Q (coding parameter prediction function) Q = g (b, A) (5). Then, for the coding parameter Q of a certain block, the encoder calculates the coding parameter prediction function Q from the code amount b allowed to be allocated to the block and the activity A of the block.
= G (b, A), the coding parameter Q is determined.

In this case, however, the biggest problem is that
The point is whether a "versatile and efficient prediction function" can be prepared.

That is, a prediction function is required that can assign codes to each block so that the encoding error of the entire image is minimized within a given amount of code regardless of the type of image. However, for pixels other than the image used to create the prediction function, the predicted values are not guaranteed. As a result, the code amount when the general image is actually coded by the coding parameter prediction function g (b, A) obtained using the sample image does not substantially match the set code amount. Therefore, generation of an undersized code that does not reach the predetermined code amount (underflow) or generation of a code exceeding the predetermined amount (overflow) always occurs.

In the former case, the image quality is lower than the image quality that may be realized, but it is not destructive.

However, in the latter case, it is necessary to perform some kind of "code truncation processing". That is, by selecting blocks whose code amount is to be reduced and changing the coding parameters of those blocks, the code amount is reduced (as a result, the image quality is inevitably reduced in those blocks), and the entire code amount is reduced. Is adjusted so that the value falls within a predetermined value.

In general, there is an experimental fact that an activity, a code amount, and a quantization error have a proportional relationship with respect to a coded image. From this, it is understood that such code amount control performs indirect control of the quantization error by using the activity as an index.

That is, this method does not necessarily directly aim at realizing the best (small quantization error) pixel which should be able to achieve the maximum with the code amount.

[0024]

As will be understood from the above description, the problems of the code amount control system using such a prediction function are summarized in the following two points.

One is that the accuracy of the prediction function is not very high, so a considerable emphasis needs to be placed on "truncation processing" to avoid failure. Therefore, the image quality varies depending on the block, and as a result, the image quality cannot be stably controlled.

The other one cannot be said to be a control which directly aims at realizing the best image quality (minimization of the quantization error).

In view of the above circumstances, the present invention obtains a (quantization error, code length) distribution of the entire image by measuring the quantization error and the code length of each block, and then, based on this distribution. Thus, as a whole image, the quantization error is the least, and the code amount is the closest to the set code amount,
It is an object of the present invention to provide a code amount control device at the time of image encoding that determines an optimal encoding parameter Q for each block.

That is, according to the present invention, the quantization error of a block is used as a direct index so that the quantization error is minimized by using the given set code amount frame as much as possible, and In order to realize coding assignment, the selection of coding parameters for each block is determined while considering the image quality and code amount of the entire image, and furthermore, achieves the set code amount stably with very high accuracy, It is an object of the present invention to provide a code amount control device at the time of image encoding, which can significantly reduce the risk of occurrence of a catastrophic overflow.

[0029]

In order to achieve the above object, a code amount control apparatus according to the present invention at the time of image encoding converts an image to be encoded into a set of N blocks. An irreversible coding method for generating a variable length code for each block, wherein a counting means for counting the length b (i, k) of the variable length code, an evaluation function calculating means for calculating an evaluation function F, And a mapping determining means for determining a mapping す る that minimizes the evaluation function F.
Quantity, Is a variable, a function for calculating the sum of them is a function E, and using a coefficient λ,

(Equation 8) Where the mapping determination means calculates the sum of code amounts,

(Equation 9) To match the code amount B0 set in advance, the blocks are selected in a predetermined order for a given function ψ, and the block selected at a certain point in time is selected as i
, The assignment of the function value ψ (i) of the block i (encoder number to be assigned to the block i) is ψ (i)
= K, the value k of the allocation to the block i is changed so as to make the value of the evaluation function F smaller, and this allocation change is sequentially performed for blocks other than i. , K constant value functions ψk of the mapping ψ
(Where k = 1,..., K) for all i
When k (i) = k, K kinds of code amount sums B obtained by performing encoder assignment by the K kinds of fixed value functions ψk
(Ψ1),..., B (ψK) and K kinds of coding error sum E
(Ψ1),..., E (ψk), and selects the code amount sum B (ψk) closest to the preset code amount B0 from the K kinds of code amount sums, and selects the selected number k Is a means for sequentially determining ψ as a constant value function ψk by as an initial value of the mapping ψ.

[0030]

In the above arrangement, k types of variable length encoders C (k) are prepared in block units, and the length of the variable length code generated when C (k) is applied to a certain block i is determined. b
(i, k), the encoding error of the block is e (i, k), and the correspondence from the subscript set {1,..., N} of the block to the subscript set {1,. When the mapping is ψ,
By minimizing the evaluation function F with the length of the 2N variable-length codes as b (i, k) and the coding error of the block as e (i, k), the sum b ( 1, ψ (1)) +… + b (i, ψ
(i)) +... + b (N, N (N)) is determined so that the mapping 割 り 当 て る to be assigned to the encoder for each block is equal to the predetermined code amount B0. The block is coded, thereby optimizing the total code amount and the sum of the quantization errors.

[0031]

【Example】

<< Basic Principle of the Present Invention >> First, while formulating an optimal code assignment problem, a control procedure of a code amount according to the present invention will be described, and then a specific configuration method of hardware for executing the procedure will be described. This will be described with reference to the drawings.

<Formulation of optimal code assignment problem and code amount control procedure according to the present invention> First, K values, {Q ₁ ,..., Q _K } (6), are prepared as coding parameters Q.

Also, the number of pixel blocks is N, and
It is assumed that the quantization error and the code length of the pixel block at the time of encoding with the encoding parameter Q _k are e (i, k) and b (i, k), respectively.

Here, assigning any one of the K types of encoding parameters {Q1,..., QK} to each of the N blocks is a set of subscripts.

(Equation 3) From the index set,

(Equation 4) Ψ, which gives a correspondence to

(Equation 5) There is no other way to determine one. That is, ψ (i)
Indicates the number of the encoding parameter to be applied to the i-th pixel block. In other words, determining one mapping ψ is equivalent to K coding parameters {Q 1,..., Q K}
, One encoding strategy for the entire image composed of N blocks has been determined.

Hereinafter, this mapping ψ is called a “sign assignment” function. Under this function ψ, the sum E (ψ) of the quantization errors of the entire image and the sum B (ψ) of the code amount are

(Equation 6) Becomes On the other hand, assuming that the predetermined total amount of codes for the entire image is B0, the problem of determining the optimal coding parameter assignment function ψ is formulated as the following constraint minimization problem [A]. Be transformed into

B (ψ) = B0 (12) [A]: “Find a function ψ that minimizes E (ψ) under the equation (12)”.
(13) By the way, a function ψ that is a solution to the above-mentioned optimization problem [A]
As is well known in Lagrange's undetermined multiplier method and the like, the following minimization problem [B], F (ψ) = E (ψ) + λ · (B (ψ) −B0) ² . (14) [B]: “Use a proper positive number λ to find a function ψ that minimizes the functional F (ψ) shown in equation (14)”.
It is obtained as the solution of (15). Then, determining the optimal coding strategy for one image in question here is to find the solution 問題 to the problem [B] for that image.

Nevertheless, obtaining an optimal solution to this optimization problem simply means searching for an optimal function ψ from among K ^N kinds of functions ψ, and if k = 2, N = 6
Even when the number is 4, the total number of candidates for the function と な り is K ^N = 1.84 × 10 ¹⁹ (16), which indicates that it is practically impossible to find an optimal solution.

However, if the method of the present invention is used, a sub-optimal function ψ can be obtained at very high speed (in real time), though not necessarily optimally.

The control procedure of the present invention for quickly finding the code assignment function ψ (accurately, a quasi-optimal solution) of the optimal solution to the problem [B] will be described below.

First, assuming that a constant value Q _k is selected as a coding parameter for all N blocks, as a code assignment function, 全 て_k (i) = k... For all i. (17) This means that the constant value function ψ _k defined as follows is used.

Then, for the input coded image,
Applying K different value functions, ψ ₁ ,..., Ψ _K (18), K values, B (ψ ₁ ),..., B (ψ _K ). can get.

From these, a constant value function that realizes the code amount closest to the set code amount B0 is selected. This is the function ψ
_{If k} , then for all N blocks,
When the same coding parameter Q _k is selected, the set code amount B
This means that the code amount B (ψ _k ) closest to 0 is generated.

Here, this constant value function ψ _k is adopted as the initial value ψ ⁽⁰⁾ of the control procedure for finding the optimum function ψ.

Next, with respect to each of the N blocks, the function ψ is updated by successively selecting an encoding parameter that minimizes the evaluation function F (ψ). .

That is, at the t-th step in this successive updating procedure, the code assignment function at this time is ψ ^(t−1)
If a block p is selected, then as a new coding parameter number k ^* for the block p, for all k, F (ψ, p, k ^* ) ≦ F (ψ, p, k) ... Defined as satisfying (20). Where B (ψ, p, k) = B (ψ) −b (p, ψ (p)) + b (p, k) (21) E (ψ, p, k) = E (ψ) −e (P, ψ (p)) + e (p, k) (22) F (ψ, p, k) = E (ψ, p, k) + λ · (B (ψ, p, k) −B0) ² ... (23) where B (ψ, p, k), E (ψ, p, k), F (ψ, p, k)
Represents the total code amount, the total quantization error, and the evaluation function when the coding parameter Q _k is applied only to the p-th block and the other blocks use the code assignment function ψ. I have.

Using the number k ^* of the encoding parameters thus selected, a new code assignment function ψ (t) is

(Equation 7) Is defined.

That is, a new ψ at a certain time t
^(t) In preparing is focused on the block p, to be smaller than the evaluation value F for the old ^{ψ (t-1) (ψ} (t-1)), look for new code assignment for block p ,
That's what it means.

In this updating method, a potential plane defined on a plane is traversed by k ^N lattice points (each lattice point corresponds to a different function ψ) on the plane while moving toward a valley. It can be interpreted figuratively as the "steepest descent method" that descends.

The steepest descent method generally does not guarantee an optimal solution for any potential function. However, as is clear from the above description, as long as this control procedure is followed, a sequence of functions さ せ る that monotonously reduces the evaluation function F is generated. It is for this reason that we have stated earlier that this control procedure gives a suboptimal solution of the minimization problem [B].

<Extension of Error Evaluation Function E (ψ)> In the explanations so far, (10)
The sum of the quantization errors (the sum of the square errors of each block) of the entire image represented by the equation has been used.

However, it is known that the evaluation of the image quality of an encoded image is not always sufficient only by the sum of the square errors. For example, in a boundary portion or an edge portion, even if the square error is small, it is perceived as a great deterioration by human eyes, and a quantization error in a fine texture portion is not so noticeable.

Therefore, it is conceivable to define a new error for each block error, e (i, ψ (i)) (25), taking into account the distribution of blocks around the block.

Also in this case, a new error evaluation function E
(Ψ) is the error of each block, e (1, ψ (1)),..., E (N,) (N))... (26) Therefore, in this case, it is clear from the expressions (20) to (24) that the code amount control procedure of the present invention can be applied without any change.

<General versatility of the present invention> First, the image coding method is defined as orthogonal transform (DCT, etc.) + Scalar quantization + variable length code (entropy code)
The problem of fixed-length code control has been described by taking (27) as an example. However, in the method of the present invention, since an amount unique to orthogonal transform such as “activity” is not used, the range of code systems that can be used is Can be significantly expanded.

For example, LOT (Lapped Orthogonal Tran)
form: orthogonal transformation in which pixel blocks to be transformed overlap each other). In the case of LOT,
The resulting transform coefficients are a completely separated set of blocks (although of a different size than the original pixel block)
Further, the information amount of the set of the entire transform coefficient block matches the information amount of the entire original image. That is, LOT
, Scalar quantization and variable-length coding are performed on each of the independent LOT transform coefficient blocks. Therefore, “block” in the above control procedure is replaced with “LOT coefficient block”.

As another example, the present invention is also applicable to encoding in which an image is scalar-quantized in block units, then vector-quantized, and the centroid number is converted into a variable-length code.

More generally, coding (block code) for converting the information of the entire image into a set of blocks of a fixed size and then generating one variable-length code for each of the blocks. In the case where a scalar control variable is used for code amount control for each block, this control procedure can be applied as it is.

<< Example of Hardware Configuration of the Present Invention >> Next, the configuration of hardware for executing the control procedure of the present invention at high speed will be described. As described above, this control procedure converts an entire image into a set of blocks of a certain size,
The present invention can be generally applied to a general encoding method in which one variable-length code is generated for each of those blocks. Therefore, in this example, it is assumed that the components of the block are not limited to orthogonal transform coefficients such as DCT, but are merely numerical values.

Then, K kinds of precoders for one block are prepared, and {C ₁ ,..., C _K } (28). This encoder generates one variable-length code for each block. Generally, different types of encoders generate different code amounts and quantization errors. Therefore, for example, as described above, a combination of a scalar quantizer and an entropy encoder or a vector quantizer that generates a centroid number as a variable-length code may be considered.

The precoder includes a local decoder, and also calculates a quantization error for each block.

FIG. 1 is a block diagram showing an embodiment of an encoding circuit under such conditions and when K = 4.

The encoding circuit shown in this figure has four precoders 1 and four quantization error / code amount map circuits 2
A code initial value selection circuit 3, a quantization error sum register circuit 4, a code amount sum register circuit 5, an optimum encoder number map circuit 6, a block number generation circuit 7, an optimum code selection circuit 8, , Delay circuit 9 and changeover switch 10
And four post-coders 11. The coding parameters Q _k for each block are minimized so that the total sum of the quantization errors is minimized under a predetermined code amount B0. Is determined and each block is encoded.

Each precoder 1 includes a local decoder, and has different coding parameters Q _k from each other.
(Where k = 1,..., 4) are set, and each time N blocks constituting the input image are supplied, this block is blocked based on the set encoding parameter Q _k. Each block is coded as a block of number i (where i is one of 1,..., N), and the code amount b (i, k) of the four output codes and the quantization error e (i,
k) are generated and supplied to the respective quantization error / code amount map circuits 2. Then, when the encoding process for the N blocks is completed, four code amount sums B _k indicating the entire code amount for the N blocks constituting the input image for each precoder 1; The sum of four quantization errors E _k indicating the entire quantization error is calculated and supplied to the code initial value selection circuit 3.

Each quantization error / code amount map circuit 2 is a RAM having N cells corresponding to the number N of blocks.
And each of the precoders 1 to 4
Code amount b (i, k) and quantization error e of four output codes
Each time (i, k) is output, it is fetched and stored in each cell of the corresponding block number i, and when a read command is output from the optimum code selection circuit 8, the read command Four code amounts b (p, k) stored in each cell corresponding to the designated block number p (where p is one of 1,..., N) and the quantization error of the four output codes e (p, k) is supplied to the optimum code selecting circuit 8.

The code initial value selection circuit 3 receives the four code amount sums B _k and the four quantization error sums E _k from the respective pre-encoders 1 when they are supplied, and also takes them in. , and the set code amount B0 which is set for the input image, of the respective code amount sum B _k by comparing each of said respective code amount sum B _k, the closest code amount sum to the set code amount B0 Is supplied to the optimal encoder number map circuit 6, and one code amount total Bk corresponding to the encoder number k is _stored in the code amount total register circuit. 5 and supplies one quantization error sum E _k corresponding to the encoder number _k to the quantization error sum register circuit 4.

When one code amount sum _Bk is output from the code initial value selection circuit 3, the code amount sum register circuit 5 fetches and stores the code amount sum Bk. When output, the stored code amount sum B _k is supplied to the optimum code selection circuit 8. When a write command is output from the optimum code selection circuit 8, the code amount total B _k output together with the write command is fetched and the code amount total B _k up to that time is taken.
Storing it in place, and supplies the optimum code selection circuit 8 after this when the read command is outputted, the code amount sum B _k stored in said optimum code selection circuit 8.

The quantization error sum register circuit 4
One quantization error sum E from the code initial value selection circuit 3
_{When k} is output, it is fetched and stored, and when a read command is output from the optimum code selection circuit 8, the stored quantization error sum E _k is supplied to the optimum code selection circuit 8. . When a write command is output from the optimum code selection circuit 8, the quantization error sum E _k output together with the write command is fetched and stored in place of the previous quantization error sum E _k , when the read command is outputted from the optimum code selection circuit 8 Thereafter, supplies the quantization error summation E _k stored in said optimum code selection circuit 8.

Also, the optimal encoder number map circuit 6
One encoder which is constituted by a RAM having N cells corresponding to the number of blocks constituting the input image and which indicates the code amount sum closest to the set code amount B0 from the code initial value selection circuit 3 When the number k is output, it is fetched and written into all cells. When a read command is output from the optimum code selection circuit 8, the encoder number k stored in the cell corresponding to the block number p specified by the read command is read, and this is used as the initial encoder. The number OPT (p) is supplied to the optimum code selection circuit 8, and when a write command is output from the optimum code selection circuit 8, the optimum coder number k ^* output together with the write command is optimized. Chemistry number OP
The data is stored in the cell corresponding to the block number p specified as T (p). When the writing of the optimal encoder numbers OPT (p) to all the cells is completed, the optimal encoder numbers OPT (i) are sequentially read from the first cell and supplied to the changeover switch 10 as a switching signal. Then, after reading out the optimal encoder number OPT (N) stored in the last cell and supplying it to the changeover switch 10 as a changeover signal, the operation is terminated.

Further, the block number generation circuit 7 supplies the first code amount sum B _{k to the} code amount sum register circuit 5.
After completion of the storage processing of the above, the storage processing of the first quantization error sum E _{k in} the quantization error sum register circuit 4, and the storage processing of the encoder number k in the optimal encoder number map circuit 6, the operation is terminated. At the start, N block numbers p are randomly generated and supplied to the optimum code selection circuit 8. Then, when the generation of all the block numbers p is completed, the operation is stopped.

Each time the block number p is output from the block number generation circuit 7, the optimum code selection circuit 8 outputs the set code amount B0 and the code amount sum B _k stored in the code amount sum register circuit 5. , fetches the quantization error summation E _k stored in the quantization error summation register circuit 4, the accesses to the respective quantization error and code amount map circuit 2 on the basis of the block number p block number p , The four code amounts b (p, k) and the four quantization errors e (p, k) stored in each cell corresponding to wherein after capturing the initial encoder number OPT (p) stored in the cell corresponding to the block number p, these settings code amount B0, code amount sum B _k, the quantization error summation E _k, the code amount Te b (p, k), each The following equation is performed on the basis of the encoding error e (p, k) and the initial encoder number OPT (p), and the optimum encoder number that satisfies the following equation is obtained for all the encoder numbers k. Find k ^* .

F (p, k ^* ) ≦ F (p, k) (2)
9) where F (p, k) = E (p, k) + λ · (B (p, k) −B0) ² (30) E (p, k) = E _k −e (p, OPT ( p)) + e (p, k) ... (31) B (p, k) = B k -b (p, OPT (p)) + b (p, k) ... (32) then, the optimum code selection circuit 8 Is the optimal encoder number k ^*
, A new sum of quantization errors E _k ^* and a sum of code amounts B _k ^* are obtained by performing the operation shown in the following equation.

[0072] ^{_{E k * = E k -e (}} p, OPT (p)) + e (p, k *) ... (33) B k * = B k -b (p, OPT (p)) + b (p, k ^* ) (34) Then, the quantization error sum E _k ^* obtained by the equation (33) is stored in the quantization error sum register circuit 4 as a new quantization error sum E _k , and the (3)
The code amount sum B _k ^* obtained by equation (4) is stored in the code amount sum register circuit 5 as a new code amount sum B _k .

Next, the optimal encoder number k ^* obtained by the above equation (29) is supplied to the optimal encoder number map circuit 6 as the optimal encoder number OPT (p), and Instead of the initial encoder number OPT (p) stored in the corresponding cell, this optimal encoder number O
PT (p) is stored.

Further, the delay circuit 9 is constituted by a FIFO memory which, each time the N blocks constituting the input image are supplied, successively delays them while taking them in and outputs them in a first-in first-out format. Each time the block is supplied, the block is fetched and sequentially delayed, and the optimum coder number OPT is assigned to all cells in the optimum coder number map circuit 6.
(p) is stored, and this optimal encoder number map circuit 6 is stored.
When the first switching signal is output from the first block, the block corresponding to the first block number 1 is output and supplied to the changeover switch 10. Thereafter, each time a switching signal is output from the optimal encoder number map circuit 6, the next block is sequentially output and supplied to the changeover switch 10.

Each time a switch signal is output from the optimum encoder number map circuit 6, the switch 10 selects the post-coder 11 specified by the switch signal, It receives the output block and supplies it to the selected post-coder 11.

Each post-encoder 11 is set with an encoding parameter Q _k corresponding to each of the pre-encoders 1, and blocks constituting an input image are supplied from the changeover switch 10. Each time, this block is encoded based on the set encoding parameter _Qk , and this is output as an output code.

Next, the encoding operation of the encoding circuit will be described with reference to FIG.

First, every time one of the N blocks constituting the input image is input, the block is assigned a block number i based on a coding parameter Q _k preset by each precoder 1. To generate a code amount b (i, k) of four output codes and a quantization error e (i, k) of the four output codes. The data is supplied to the map circuit 2 and stored in each cell of the corresponding block number i.

When the encoding process for the N blocks is completed by each pre-encoder 1, four code amount sums B _k indicating the total code amount for the N blocks constituting the input image. And the four quantization error sums E _k indicating the entire quantization error are calculated and supplied to the code initial value selection circuit 3.

Thus, the set code amount B0 set for the input image by the code initial value selection circuit 3
When, among the code amount sum B _k the four code amount sum B _k output from the pre-encoder 1 is compared each one indicating the closest code amount sum to the set code amount B0 An encoder number k is selected, supplied to the optimal encoder number map circuit 6, and stored in all cells corresponding to all blocks, and one code amount corresponding to the encoder number k is selected. The sum B _k is supplied to and stored in the code amount sum register circuit 5, and one quantization error sum E _k corresponding to the encoder number k is supplied to the quantization error sum register circuit 4 and stored. .

Thereafter, the block number generation circuit 7 starts operating, and one of the N block numbers p is randomly generated by the block number generation circuit 7 and supplied to the optimum code selection circuit 8.

When the first block number p is output from the block number generation circuit 7, one code stored in the set code amount B 0 and the code amount sum register circuit 5 by the optimum code selection circuit 8. Sum B
_k , one quantization error sum E _k stored in the quantization error sum register circuit 4 is taken in,
Each of the quantization error / code amount map circuits 2 is accessed based on the block number p, and four code amounts b stored in four cells corresponding to the block number p.
(p, k) and four quantization errors e (p, k) are fetched,
Further, after the optimum encoder number map circuit 6 is accessed and the initial encoder number OPT (p) stored in one cell corresponding to the block number p is fetched, the set code amount B0, Code amount sum B _k , quantization error sum E _k , code amount b (p, k), quantization error e (p, k)
, Based on the initial encoder number OPT (p).
Equations (9) to (32) are calculated, and the optimum encoder number k that satisfies the equation (29) is calculated for all encoder numbers k.
^* Is required.

[0083] After this, the based on the optimum coder number k ^* by the optimum code selection circuit 8 (33), (3
4) The calculation shown in the equation is performed to obtain a new code amount sum B.
_k ^* and a new quantization error sum E _k ^* are obtained.

Then, the quantization error sum E _k ^* is stored in the quantization error sum register circuit 4 as a new quantization error sum E _k , and the code amount sum B _k ^* is stored in the new code amount sum B _k ^. _The code amount is stored in the code amount sum register circuit 5 as _k .

Next, the optimal encoder number k ^* obtained by the above equation (29) is supplied to the optimal encoder number map circuit 6 as the optimal encoder number OPT (p), and This optimal encoder number OPT (p) is stored instead of the initial encoder number OPT (p) stored in the corresponding cell.

Thereafter, each time a new block number p is generated by the block number generating circuit 7, the above-described operation is repeated and the initial encoding stored in the corresponding cell of the optimal encoder number map circuit 6 is performed. The unit number OPT (p) is rewritten to the optimal encoder number OPT (p).

The initial encoder numbers OPT stored in all cells of the optimal encoder number map circuit 6
When (p) is rewritten to the optimal encoder number OPT (p), the optimal encoder number OPT (i) stored in the first cell is read out by the optimal encoder number map circuit 6. This is supplied to the changeover switch 10 as a changeover signal, and the post encoder 11 corresponding to the changeover signal is selected.

In parallel with this operation, every time the N blocks constituting the input image are supplied by the delay circuit 9, they are taken in and sequentially delayed, and the optimum encoder number map circuit is provided. When the first switching signal is output from 6, the block corresponding to the first block number 1 is output and supplied to the changeover switch 10,
Encoding is performed by the post-encoder 11 selected by the changeover switch 10, and a code obtained by this encoding operation is output as an output code.

Hereinafter, the optimal encoder number OPT (i) stored in the next cell from the optimal encoder number map circuit 6
Is read out, and this is used as a changeover signal as the changeover switch 1
0, and every time the post-encoder 11 is selected by the changeover switch 10, the next block is output from the delay circuit 9, and the next block is selected by the changeover switch 10. And output as an output code.

Then, the optimal encoder number OPT stored in the last cell from the optimal encoder number map circuit 6 is obtained.
(N) is read out and supplied to the changeover switch 10 as a changeover signal. The changeover switch 10 selects the post-encoder 11 and outputs the last block from the delay circuit 9. Selector switch 10
After the encoding is performed by the post-encoder 11 selected and output as an output code, the encoding operation of the encoding circuit ends.

[0091]

As described above, the code amount control according to the present invention is highly versatile. After converting the information of the entire image into a set of blocks of a fixed size, the variable length control is performed for each of the blocks. Any coding method (block coding) that generates one code can be applied.

Also, unlike the conventional code amount control method,
Since the quantization error of the encoded image is directly controlled, it is possible to realize the best image quality within a predetermined code amount range.

Further, the larger the number N of blocks is, the higher the probability of realizing the optimum code assignment becomes. Therefore, the present invention is suitable for encoding a large-scale image such as a high-vision image.
For this reason, DCT and LO for HDTV images
By encoding T (both in the case of generating an 8 × 8 transform coefficient block), the error of the achieved code amount can be suppressed to 1/1000 or less.

As a result, the specific gravity of the code truncation process for avoiding a failure due to overflow is significantly reduced, and a stable encoded image quality can be obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of a code amount control device at the time of image encoding according to the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Pre-encoder 2 Quantization error / code amount map circuit 3 Code initial value selection circuit 4 Quantization error sum register circuit 5 Code amount sum register circuit 6 Optimal encoder number map circuit 7 Block number generation circuit 8 Optimal code Selection circuit 9 Delay circuit 10 Changeover switch 11 Post encoder

Claims (1)

(57) [Claims] An irreversible coding method for converting an image to be coded into a set of N blocks, and then generating a variable length code for each block, comprising: counting means for counting the length b (i, k) of a variable-length code generated when the variable-length encoder C (k) is applied to a certain block i; A coding error calculating means for calculating a coding error e (i, k) of a block, and a correspondence between a subscript set {1,..., N} of a block to a subscript set {1,. When the map to be given is 、 ２, 2N quantities, Evaluation function calculation means for calculating an evaluation function F based on: and mapping determination means for determining a mapping す る that minimizes the evaluation function F, wherein the evaluation function calculation means comprises N quantities, Is a variable, a function for calculating the sum of them is a function E, and a coefficient λ is used to obtain The mapping determination means calculates the sum of the code amounts, Is set to a predetermined code amount B0, the blocks are selected in a predetermined order with respect to a function 、 given in advance, and when a block selected at a certain time is defined as i, the block i When the allocation of the function value ψ (i) (encoder number to be allocated to the block i) is ψ (i) = k, the allocation of the evaluation function F to the block i is made smaller so that the value of the evaluation function F becomes smaller. The value k is changed, and this allocation change is sequentially performed on blocks other than i, and K kinds of fixed-value functions ψk (where k = 1,
.., K) for all i, ψk (i) = k, and K kinds of code amount sums B (ψ) obtained by performing encoder assignment by the K kinds of constant value functions ψk.
1),..., B (ψK) and K kinds of coding error sum E (ψ
1),..., E (ψk), and a predetermined code amount B from the K total code amounts.
Means for selecting a code amount sum B (ψk) closest to 0 and sequentially determining ψ as an initial value of a mapping を with a fixed-value function に よ る k according to the selected number k. Code amount control device at the time.