JPH08280015A - Image encoder - Google Patents

Abstract

PURPOSE: To enable efficient encoding by accurately estimating a quantization coefficient SF generating any positive error. CONSTITUTION: Converting processing at an image converting part 1, quantizing processing at a quantizing part 2 with quantizing width from a quantizing width estimating part 10, encoding processing at a variable length encoding part 3 and the amount of codes at a code amount calculating part 4 are repeatedly calculated. An inclination calculating part 8 calculates the inclination of an approximate straight line by selecting two sets of quantization coefficients and code amounts larger or smaller than the target quantization coefficient when the shape of a graph showing the logarithm of a quantization coefficient and the logarithm of a code amount is convex upward, and by selecting two sets of coefficients with the target amount in between when the graph is convex downward. A quantization width calculation part 9 calculates the quantization coefficient from the approximate straight line and sends the quantization width to the quantizing part 2. The approximate straight line becomes always made larger than the graph near the target quantization coefficient and the error of the code amount becomes always negative.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an image coding apparatus for performing lossy coding processing, and more particularly to an image coding apparatus for controlling a code amount.

[0002]

2. Description of the Related Art Image data encoding methods are roughly classified into two types: variable length encoding methods and fixed length encoding methods. In the former case, generally, the amount of code generated cannot be known unless encoding is performed. Therefore, when the allowable code amount has an upper limit, it is necessary to control the code amount by some means. Therefore, for example, Japanese Patent Application Laid-Open No. 4-2291 discloses an image data encoding apparatus and an encoding method for controlling the code amount and compressing the image data until the image data fits in the target code amount.

FIG. 2 is a block diagram showing an example of an image coding apparatus using a conventional variable length coding system. In the figure,
1 is an image conversion unit, 2 is a quantization unit, 3 is a variable length coding unit,
4 is a code amount calculation unit, 5 is a control unit, 10 is a quantization width estimation unit, 11 is image data, 12 is conversion data, 13 is quantization data, 14 is code data, 15 is partial code amount data, 16, 23 is total code amount data, 17 is control data,
22 is quantization width data.

The image conversion unit 1 receives the input image data 11
The conversion process is performed on. The transform process is a process for compressing image information, such as orthogonal transform or estimated coding. The data converted by the image converter 1 is sent to the quantizer 2 as converted data 12. The quantization unit 2 is based on the quantization width data 22 instructed by the quantization width estimation unit 10,
The input conversion data 12 is quantized. Then, the result of the quantization processing is sent to the variable length coding unit 3 as the quantized data 13. The variable length coding unit 3
A variable length coding process is performed on the quantized data 13. As a result, depending on whether the instruction from the control unit 5 is the statistical processing or the encoding processing, the partial code amount data 15 is sent to the code amount calculation unit 4 in the statistical processing, and the code data 1 is used in the encoding processing.
4 is output to the outside. The code amount calculation unit 4 integrates the partial code amount data 15 to calculate the total code amount in units of the entire image, color components and the like. Calculated total code amount data 1
6 and 23 are sent to the control unit 5 and the quantization width estimation unit 10, respectively. The control unit 5 compares the total code amount data 16 with the target code amount designated in advance, and instructs the encoding process if the difference is within an appropriate allowable range, and instructs the statistical process otherwise. To do. The instruction is sent to the quantization width estimation unit 10 as the control data 17 together with the information on the target code amount. The quantization width estimation unit 10 estimates the quantization width for obtaining the target code amount based on the control data 17 and the total code amount data 23, and sends the result to the quantization unit 2 as the quantization width data 22. Give instructions.

3 and 4 are flowcharts showing an example of the operation in the image coding apparatus using the conventional variable length coding system. In FIG. 3, in S31, an allowable code amount is set for the control unit 5. In S32,
Image data 11 is input to the image conversion unit 1. S33
Then, a series of encoding processing is performed on the image data 11.

Details of the encoding process in S33 will be described with reference to FIG. In S41, image conversion processing relating to image information compression such as orthogonal transformation and estimation processing is performed on the image data 11 input in the image conversion unit 1. In S42, the quantization unit 2 performs quantization processing on the conversion data 12 converted in S41 using the quantization width data 22 instructed by the quantization width estimation unit 10. In S43, in the variable length coding unit 3,
Variable length coding processing is performed on the quantization width data 22.

Returning to FIG. 3, in S34, it is determined whether the instruction from the control unit 5 is statistical processing or coding processing. If statistical processing, the process proceeds to S35, and if encoding processing, the process proceeds to S40. When the process proceeds to S40, the code data 14 is output to the outside and the process ends.

On the other hand, in S35, the code amount calculating section 4 calculates the total code amount in the unit of the entire screen or the color component from the integrated partial code amount data 15. In S36, the control unit 5 determines whether or not to repeat the iterative process from S32 to S38. For example, when the termination condition is that the code amount falls within the allowable range, it is determined whether or not the difference between the total code amount data 16 and the target code amount set in S31 is included in the set allowable range. . If it is not included in the allowable range, the process proceeds to S37, and if it is included in the allowable range, the process proceeds to S39. The above-mentioned Japanese Patent Laid-Open No. 4-2291 also describes an example in which the number of iterations is used as the termination condition as the determination criterion in S36.

When the process proceeds to S39, the control unit 5 instructs the encoding process, and the process returns to S33. When the process proceeds to S37, the control unit 5 instructs the statistical processing. And S
38, the quantization width estimation unit 10 controls the control data 17
Based on the total code amount data 23, the quantization width is estimated by a predetermined method. Then, after the estimation process ends, the process returns to S33.

In the above processing, the quantization processing of S42 is performed based on the quantization width data 22 from the quantization width estimation unit 10. The conversion data 12 obtained by the image conversion processing in S41 is constant regardless of the number of iterations, and the amount of code data obtained by the variable length encoding processing in S43 depends on the amount of quantized data 13. That is, by controlling the quantization width data 22, the quantization data 13 can be controlled, and the total code amount of the code data 14 finally obtained can be controlled.

The quantization width estimation process will be described below. The above-mentioned Japanese Patent Application Laid-Open No. 4-2291 describes the following points as an example. Looking at the relationship between the relative ratio to the quantization width (hereinafter, referred to as a quantization coefficient) SF and the code amount CL for a certain image data, the following formula is established.

[Equation 1] In the equation (1), a is a constant that does not depend on the image, and b is a constant that depends on the image. Therefore (1)
If a and b can be determined to be appropriate values in the equation, the quantization coefficient SF for obtaining a desired code amount can be approximately obtained. The initial value of the quantization width is calculated from the value obtained here, and thereafter, the quantization coefficient SF corresponding to the desired code amount CL can be obtained by using a method such as the Newton-Raphson method. it can.

In addition, JPE, which is a standard method of image coding,
G (Joint Photographic Expe
rts Group) also uses the same formula. However, here, the parameter a is also dependent on the image like b.

FIG. 5 is a graph showing an example of the relationship between the quantized coefficient SF and the code amount CL obtained from the experiment. When the relationship between the quantized coefficient and the code amount is examined for an actual image, as shown in FIG. 5, a graph of the logarithm of the quantized coefficient SF and the logarithm of the code amount CL (hereinafter may be simply referred to as a graph). Is not a strict straight line, and as a whole, the quantized coefficient SF is small and is convex upward in a portion with a large code amount CL, and is convex downward in a portion with a large quantized coefficient SF and a small code amount CL. Becomes In addition, looking at this graph in detail, it can be said that it is not necessarily flat, because there are many small irregularities. For this reason, there is a difference between the assumption using the above equation (1) and the actual experimental result.

FIG. 6 shows the logarithm of the quantization coefficient SF and the code amount C.
It is explanatory drawing at the time of seeing the logarithmic graph of L microscopically.
The quantization coefficient SF takes a continuous value, but the code amount CL can take only a discrete value due to the characteristic of the processing called quantization.
Therefore, when the graph shown in FIG. 5 is enlarged, it is considered that the graph has a step shape as shown in FIG. For this reason as well, it is considered that there is a difference between the assumption using the equation (1) and the actual experimental result.

FIG. 7 is a graph showing variations in the value of the inclination a due to differences in images. FIG. 7 shows JPEG Base for 10 types of images with chart numbers 0-9.
The code | symbol amount CL with respect to some quantized coefficient SF is measured using the line method, and the result which calculated | required the parameter a of Formula (1) by multiple regression analysis is shown. Note that in FIG. 7, each image is decomposed into three components in the Lab space, and the code amount CL for each component is obtained. In the figure, ∘ indicates the value of the parameter a calculated from the code amount CL of the L component, similarly, ∘ indicates the value of the parameter a of the a component, and x indicates the value of the parameter a of the b component. As can be seen from FIG. 7, the parameter a in the above equation (1) cannot be said to be constant regardless of the image.

As described above, there is a difference between the assumption using the above equation (1) and the actual experimental result. This difference causes some problems. Particularly, according to the above-mentioned Japanese Patent Laid-Open No. 4-2291, Newton
Although the Rapson method is used for the estimation process, if the Newton-Raphson method is used as it is, the estimation process may not work well.

FIG. 8 is an explanatory diagram of the principle of the Newton-Raphson method. Now, the function f (x) exists, and some Y
Suppose we want to find X such that Y = f (X). The Newton-Raphson method is performed in the following procedure. First, an appropriate initial value x ₀ is selected and f (x ₀ ) is calculated. Next, the differential coefficient of f (x) at x ₀ , f ′
Find (x ₀ ). Then, x ₁ is calculated by the following formula.

[Equation 2] When x ₁ is obtained, the same process is repeated using x ₁ instead of x ₀ to obtain x ₂ . While this is repeated, x _n (n = 0, 1, 2, ...) Asymptotically approaches X, so the process ends when the difference between Y and f (x _n ) becomes an allowable error. .

In the above-mentioned Japanese Patent Application Laid-Open No. 4-2291, the procedure of the Newton-Raphson method is used to consider, for example, logSF as x and logCL as f (x) to obtain an arbitrary code amount CL. The quantization coefficient SF is determined. However, in practice, the Newton-Raphson method cannot be used directly to obtain the quantized coefficient SF. Because in Newton-Raphson method, f (x)
This is because f (x) can be applied only to a differentiable function because the differential coefficient f ′ (x) of is used. As described with reference to FIG. 6, the graph of the logarithm of the quantized coefficient SF and the logarithm of the code amount CL is considered to be stepwise. Such a step-like function cannot be differentiated at the round of steps, in other words, the differential coefficient becomes infinite.

Japanese Unexamined Patent Publication No. 4-2291 does not describe a method for avoiding this problem. If, as described in this document, the parameter a is constant regardless of the image, it is not necessary to obtain the parameter a, and therefore the Newton-Raphson method itself need not be used. However, as shown in FIG.
The value of is not constant and such an assumption is unreasonable. In conclusion, the parameter a should be calculated according to the image, and the Newton-Raphson method needs to be extended and applied.

Therefore, consider expanding the above equation (2) discretely so that differentiation is unnecessary. According to this, it becomes possible to perform estimation by the equation (3).

(Equation 3) Here, Δx is a minute number. Since the limit of Δx → 0 in the expression (3) is equal to the expression (2), it can be easily understood that the expression (3) is an approximation of the expression (2). If the discrete Newton-Raphson method according to the equation (3) is used, it is possible in principle to obtain the quantization coefficient SF that gives the desired code amount CL.

Further, expanding the interpretation of Δx in the equation (3), it is conceivable to use x _j (j ≠ i) instead of x _i + Δx. Therefore, the result of the previous prediction for x _j ,
That is, let's use x _i-1 . Equation (3) can be expressed as follows.

[Equation 4] That is, the slope is calculated from the i-th result and the (i-1) -th result. Since the value deviating from the definition of Δx is substituted into the equation (3), the calculation accuracy may decrease, but as in the equation (3), x _i and (x _i + Δx)
The processing load can be reduced as compared with an equation that requires data of two points.

However, these approaches give rise to some problems. This problem is caused by an estimation error due to the presence of unevenness in the entire graph as shown in FIG. FIG. 9 is an explanatory diagram of the estimation error when the graph has an upwardly convex curve. Now, the quantization coefficient SF
_The next quantized coefficient SF ₂ is estimated using ₀ and SF ₁ .
In this case, in FIG. 9, the alternate long and short dash line connecting the quantization coefficients SF ₀ and SF ₁ corresponds to the straight line of the above-mentioned expression (1). Therefore, the quantization coefficient SF ₂ is obtained as a point on the straight line that corresponds to the target code amount CL. At this time, since this straight line is not the same as the actual graph, when the SF ₂ is actually encoded, an error as shown in FIG. 9 occurs.

In this example, the quantization coefficients SF ₀ and SF
₁ exists on both sides of the desired quantization coefficient SF. In such a case, as shown by the alternate long and short dash line in FIG. 9, the curve of the graph always exists above the straight line of the equation (1) at the quantization coefficient SF ₂ . Therefore, the error is the error toward the one with the larger actual code amount. In the code amount control, since the code amount is often suppressed to a certain code amount or less, it is not desirable that the code amount CL exceeds the target value.

FIG. 10 is an explanatory diagram of the estimation error when the graph has a downward convex curve. In FIG. 10, two quantized coefficients SF ₀ and SF ₁ exist on the side larger than the desired quantized coefficient SF. At this time, the alternate long and short dash line connecting the quantization coefficients SF ₀ and SF ₁ corresponds to the straight line of the equation (1). Therefore, the quantization coefficient SF ₂ is obtained as a point on the straight line that corresponds to the target code amount CL. However, as in the case where the graph is convex upward, the quantization coefficient SF
_{In 2} , the curve of the graph is always above the straight line in equation (1). Therefore, an error occurs in the side where the actual code amount is large. Also, when both of the quantized coefficients SF ₀ and SF ₁ are present on the side smaller than the desired quantized coefficient SF, an error occurs in the direction in which the actual code amount is large for the same reason.

As described above, in the conventional estimation process,
There is a problem that an error may occur in a direction in which the actual code amount is larger than the target code amount.

[0026]

SUMMARY OF THE INVENTION The present invention has been made in view of the above-mentioned circumstances, and it is possible to accurately estimate the quantization coefficient SF without causing a positive error and perform efficient encoding. It is an object of the present invention to provide an image encoding device capable of performing the above.

[0027]

According to the present invention, in an image coding apparatus, image conversion means for converting an image and outputting converted data, and quantization according to a quantization width given to the converted data. Quantization means for outputting quantized data, variable length coding means for performing variable length coding on the quantized data, and unit image by integrating the code amount created by the variable length coding means And a control means for comparing the code amount of the unit image created by the code amount calculating means with a preset target code amount and performing control according to the result. Under the control of the control means, the quantized coefficient is calculated based on the shape of the graph of the logarithm of the quantized coefficient and the logarithm of the corresponding code amount and the target code amount obtained from the encoding results up to that point, and further the quantized Based on coefficient It is characterized in that it has a quantization width estimating means calculates the quantization width supplied to the quantization means are.

When the shape of the graph of the logarithm of the code amount corresponding to the logarithm of the quantization coefficient is convex upward, the quantizing width estimating means sets the target code amount. The two coding results that give the larger code amount and the code amount closest to the target code amount, or the two coding results that give the code amount smaller than the target code amount and closest to the target code amount are selected, It can be configured to calculate a quantization factor and the quantization width.

Further, as in the invention described in claim 3, when the shape of the graph of the logarithm of the code amount corresponding to the logarithm of the quantization coefficient is convex downward, One encoding result that gives a code amount that is greater than the code amount and closest to the target code amount and one encoding result that gives a code amount that is less than the target code amount and closest to the target code amount are selected and It can be configured to calculate a quantization factor and the quantization width.

Further, as in the invention described in claim 4,
The quantization width estimation means is a parameter selection means for adaptively selecting a parameter creation method according to the control from the control means, and a slope parameter and a quantization which are predetermined constants according to the control from the parameter selection means. Constant presenting means provided as an initial value of the coefficient, slope calculating means for calculating the slope parameter based on the shape of the graph of the logarithm of the quantization coefficient corresponding to the logarithm of the quantization coefficient according to the control from the parameter selecting means, It may be configured to have a quantization width calculation means for calculating the quantization coefficient and the quantization width based on the slope parameter given from the constant presenting means or the slope calculating means and the target code amount.

According to the invention of claim 5, the constant presenting means in the image coding device of claim 4 is as follows.
The absolute value of the initial value of the slope parameter to be given to the quantization width calculation means is larger than the expected value when the graph of the logarithm of the code amount corresponding to the logarithm of the quantization coefficient is convex upward, Further, when the graph is convex downward, it can be configured to be set smaller than an expected value.

At this time, as in the invention described in claim 6, the constant presenting means sets the initial value of the quantization coefficient given to the quantization width calculating means to an average value obtained in advance. Can be configured.

Further, as in the invention described in claim 7,
The constant presenting means may be configured to prepare a plurality of parameters to be given to the quantization width calculating means and selectively use the parameters according to the target code amount.

In the image coding apparatus according to any one of claims 1 to 7 described above, the control means, as in the invention described in claim 8, has the code amount of the unit image and the target code amount. The estimation process and the encoding process may be repeated for the unit image until the difference between and reaches a certain allowable value, and then control may be performed to output the code.

Further, in the image coding apparatus according to any one of claims 1 to 8, the image conversion means includes:
According to a ninth aspect of the present invention, a conversion data storage unit that stores the conversion data is provided, and the conversion data can be repeatedly read as necessary.

Further, in the image coding apparatus according to any one of claims 1 to 9, the variable length coding means stores the code created as in the invention described in claim 10. A storage means is provided to store the latest code, a code having a code quantity closest to the target code quantity, or a code having a code quantity closest to or less than the target code quantity, so that the code can be read out as necessary. Can be configured to.

[0037]

As shown in FIG. 9 described above, in the case of an upwardly convex graph, when estimated from two points sandwiching the originally desired quantization coefficient SF, the actual graph is on a straight line connecting the two points. Since it passes, the error always appears in the direction in which the code amount increases. Further, in the downward convex graph, as shown in FIG. 10, when estimated from two points existing on either the larger side or the smaller side of the originally desired quantization coefficient SF, the actual graph is It passes over the straight line connecting the points, and the error appears in the same way as the code amount increases. On the contrary, in the graph which is convex upward, if two points do not sandwich the quantized coefficient SF that is desired to be obtained, the error appears in the smaller code amount. Even in the downward convex graph, if two points sandwiching the desired quantized coefficient SF are used, the error appears in the direction of smaller code amount. In order to make the error always minus, it is sufficient to perform the processing in consideration of such a property. Hereinafter, a set of the quantization coefficient SF and the code amount CL in the past coding result will be referred to as estimation data.

Therefore, according to the present invention, in the quantization width estimating means, the graph shape and the target code amount of the logarithm of the code amount CL corresponding to the logarithm of the quantization coefficient SF obtained from the encoding results up to that time. The quantization coefficient SF is estimated based on In particular, when the shape of the graph is convex upward, as in the invention according to claim 2, two coding results that give a code amount that is larger than the target code amount and closest to the target code amount, or the target The quantization coefficient SF is estimated by selecting two coding results that give a code amount smaller than the code amount and closest to the target code amount. When the shape of the graph is convex downward, one encoding result that gives a code amount larger than the target code amount and closest to the target code amount, and the target code, as in the invention according to claim 3, Select one coding result that gives a code amount that is less than the amount and is closest to the target code amount,
The quantization coefficient SF is estimated. By performing such an estimation, the error always becomes negative, and the actual code amount does not exceed the target value.

In the case where code data is output after two estimation processes, that is, the quantization coefficient SF is estimated from the two estimation data, the quantization width is determined and quantization is performed, and variable length encoding is performed. If it is done, further improvement is required. In this case, when the graph is convex upward, the quantization coefficient SF ₁ of the estimation data obtained in the second estimation process exceeds the quantization coefficient SF obtained from the initial value SF _{0 of the} quantization coefficient. Don't That is, the quantization coefficients SF ₀ and SF
The positional relationship of ₁ should not be the relationship shown in FIG. 9 described above. Therefore, it is necessary to increase the absolute value of the slope parameter a when obtaining the quantization coefficient SF ₁ . On the contrary, when the graph is convex downward, the absolute value of the tilt parameter a may be set to be small.

In order to realize this, according to the invention described in claim 4, in the quantization width estimating means, there is provided parameter selecting means for adaptively selecting a method for creating parameters, and control from the parameter selecting means. According to the control from the parameter selecting means, a constant presenting means for providing a predetermined constant as a slope parameter and an initial value of the quantizing coefficient, and a graph shape of a logarithm of the quantizing coefficient and a logarithm of the corresponding code amount. A slope calculation means for calculating a slope parameter based on the slope parameter; a quantization width calculation for calculating the quantization coefficient and the quantization width based on the slope parameter given by the constant presenting means or the slope calculation means and the target code amount. The estimation is started by changing the tilt parameter according to the shape of the graph. In particular, as in the invention described in claim 5, the absolute value of the initial value of the slope parameter given to the quantization width calculation means by the constant presentation means is set to be larger than the expected value when the graph is convex upward. Set to a larger value, or smaller than the expected value if the graph is convex downward. With such a configuration, it is possible to perform an estimation process such that the error does not necessarily become negative without causing the states shown in FIGS. 9 and 10 in the estimation process, and the actual code amount is the target. It is possible to perform encoding without exceeding the code amount of. At this time, as the initial value of the quantization coefficient which the constant presenting means gives to the quantization width calculating means together with the slope parameter, it is possible to use an average value obtained in advance as in the invention described in claim 6.

Further, as described with reference to FIG. 5, the relationship between the logarithm of the quantization coefficient SF and the logarithm of the code amount CL is the code amount C.
When L is small, the graph is convex downward and the code amount CL
If there are many, the graph will be convex upward. By utilizing such a property, as in the invention according to claim 7, the constant presenting means prepares a plurality of parameters to be given to the quantization width calculating means, and selectively uses them according to the target code amount. Can be configured to. As a result, no matter what target code amount is given, it is possible to select and use an appropriate initial value and perform good estimation of the quantized coefficient.

According to the eighth aspect of the present invention, the estimation processing and the encoding processing as described above are repeated until the difference between the code amount of the unit image and the target code amount reaches a permissible value, and after that, The control means controls to output the code. As a result, the error in the code amount can be kept within the allowable range. Of course, by performing the above estimation process, the error is always negative and the actual code amount does not exceed the target code amount.

Further, according to the invention described in claim 9, the image conversion means is provided with a conversion data storage means for storing the conversion data, and is constituted so that the conversion data can be repeatedly read as necessary. As a result, in the iterative estimation processing and encoding processing, since the stored conversion data only has to be read in the estimation processing and the encoding processing after the second time, processing for generating unchanged conversion data each time. Since it is not necessary to perform the above, it is possible to reduce the overall processing amount and efficiently perform the processing after the quantization processing.

According to the tenth aspect of the invention, the variable length coding means includes code storage means for storing the created code, and has the latest code or the code having the code quantity closest to the target code quantity. Alternatively, a storage means for storing the code having the closest code amount less than or equal to the target code amount is provided, and the code can be read as necessary. As a result, when the code obtained by the estimation process is the best, the stored code data may be read and output without performing the quantization process and the coding process again. Therefore, the image coding device Throughput can be improved.

[0045]

1 is a block diagram showing an embodiment of an image coding apparatus of the present invention. In the figure, the same parts as those in FIG. 6 is a parameter selection unit, 7 is a constant presentation unit, 8 is a slope calculation unit, 9 is a quantization width calculation unit, 18 and 19 are control data, and 20 and 21 are parameter data.

The parameter selection unit 6 adaptively selects a parameter creation method according to the contents of the control data 17 from the control unit 5, and instructs the constant presenting unit 7 or the slope calculating unit 8 to calculate and present the parameters. To do. The constant presentation unit 7 sends the predetermined slope a and the value of the quantization coefficient SF as parameter data 20 to the quantization width calculation unit 9. The gradient calculator 8 estimates the gradient of the graph of the relationship between the code amount CL and the quantization coefficient SF from the control data 19 given from the parameter selector 6 and sends it as parameter data 21 to the quantization width calculator 8. The quantization width calculation unit 8 calculates a quantization coefficient SF predicted to correspond to the target code amount from the given parameter data 20 or 21 and converts it into a quantization width to quantize the quantization width data 22. Conversion unit 2
Send to. At this time, when the quantization coefficient SF is directly designated by the parameter data 20, the quantization coefficient S
The calculation process of F is not performed, and the quantization width is obtained from the given quantization coefficient SF.

The conversion method used in the image conversion unit 1 is as follows.
For example, the Discrete Cosine Transform (DCT) can be used. At this time, the transform data 12 is a DCT transform coefficient. As a variable length coding method used in the variable length coding unit 3, for example, Huffman coding can be used. Of course, other conversion methods and variable length coding methods can also be used.

Further, the conversion data 12 is stored in the image conversion unit 1.
A memory may be provided to store the. In the estimation process, the transform data 12 is repeatedly quantized while changing the quantization width, but the transform data 12 to be quantized is the same data. Therefore, by storing the converted data 12 in the memory and reading it as necessary, it is not necessary to repeatedly perform the same conversion process, and the time required for the quantizer 2 to obtain the converted data 12 Can be shortened to shorten the entire processing time.

Further, the variable length coding unit 3 can be provided with a memory for storing codes. When the current code can be output as it is, such as when the code amount of the obtained code is very close to the target code amount, the code in the memory is read and output as the code data 14. As a result, it is not necessary to perform the final quantization processing and coding processing.
Overall processing time can be shortened and throughput can be improved.

FIG. 11 is a flow chart for explaining an example of the operation in one embodiment of the image coding apparatus of the present invention. In S51, the allowable code amount is set in the control unit 5. In S52, the constant presenting unit 7 sends the value preset as the initial value of the quantization coefficient SF to the quantization width calculating unit 9.

In S53, the image data 11 is input to the image conversion unit 1. In S54, a series of encoding processing is performed on the image data 11. This encoding process
Since it is as shown in FIG. 4, the description is omitted here. However, in the quantization processing in S42, when the quantization coefficient SF is passed from the constant presenting section 7 to the quantization step calculation section 9 as the parameter data 20, the quantization step calculation section 9 uses the passed quantization coefficient. Quantization width data 22 from SF
Is generated, and the quantization processing based on this is performed by the quantization unit 2. When the parameter data 21 is passed from the slope calculator 8 to the quantization width calculator 9, the quantization width calculator 9 estimates the quantization coefficient SF and estimates the estimated quantization coefficient S.
The quantization width data 22 is created from F, and the quantization unit 2 performs the quantization processing based on the quantization width data 22.

The processing of S55 to S58, S61 and S62 is the same as the processing of S34 to S37, S39 and S40 in FIG. That is, in S55, it is determined whether the instruction from the control unit 5 is the statistical processing or the encoding processing. If the instruction is the encoding processing, the encoded data 14 is output to the outside in S62 and the processing ends. On the other hand, in the case of statistical processing, the process proceeds to S56, and the code amount calculation unit 4 calculates the total code amount in the unit of the entire screen or color component from the integrated partial code amount data 15. In S57, the control unit 5
At, it is determined whether or not to repeat the iterative process from S53 to S60. When continuing the iterative process, the control unit 5 instructs the statistical process in S58, and when ending the iterative process, the control unit 5 instructs the encoding process in S61. In either case, the process proceeds to S59, where the parameter selection unit 6
In addition, either the constant presenting unit 7 or the slope calculating unit 8 sets the value of the slope used in the quantization width calculating unit 9. Then, after estimating the quantization coefficient in S60,
Return to S53.

FIG. 12 is a flow chart showing an example of the operation of the inclination setting process. The inclination setting process performed in S59 is performed as follows, for example. In S71, it is determined whether the instruction from the control unit 5 is an encoding process or a statistical process. In the case of the encoding process, the estimation data that meets the conditions is selected in S72, and the process proceeds to S76. In the case of statistical processing, in S73, in the parameter selection unit 6, S53.
It is determined whether the number of iterations from S60 to S60 is the first time, and 1
In the case of the first time, the constant presenting unit 7 is instructed to create the parameter data. In this case, the process proceeds to S74, the value of the parameter a preset in the constant presenting unit 7 is sent to the quantization width calculating unit 9 as the parameter data 20, and S5
The process of 9 is completed. On the other hand, when the iterative process is the second time or later, the parameter selection unit 6 instructs the inclination calculation unit 8 to create parameter data. Then, the process proceeds to S75, and the latest two estimation data are selected. In S76, the slope a is estimated from the estimation data selected by the slope calculating unit 8 and sent as parameter data 21 to the quantization width calculating unit 9, and S59 ends.

The quantization coefficient estimation process in one embodiment of the image coding apparatus of the present invention will be further described below. As described above, the graph of the logarithm of the quantization coefficient SF and the logarithm of the code amount CL is, for example, as shown in FIG.
Concavities and convexities are seen instead of the straight line as shown in the above formula (1). By switching the initial values of the parameters used at the time of estimation and the method of taking the estimation parameters according to the shape of such a graph, the code amount actually obtained can be kept within the target code amount. The following is a summary of the shape of the graph, the initial values of the parameters for estimation at that time, and the method of taking them.

First, when the graph is convex upward, (1) When the estimation process is limited to two times, the absolute value of the initial value of the parameter a used in the first estimation process is assumed to be the assumed slope. Take it larger. (2) Used to find the final quantized coefficient SF 2
The two pieces of estimation data are selected from the combinations which do not sandwich the quantized coefficient SF to be obtained and which give the result closest to the target code amount.

When the graph is convex downward, (1) When the estimation process is limited to two times, the absolute value of the initial value of the parameter a used in the first estimation process is assumed to be the assumed slope. Take smaller. (2) Used to find the final quantized coefficient SF 2
The two pieces of estimation data are selected from the combinations which sandwich the quantized coefficient SF to be obtained and which give the result closest to the target code amount.

The assumed slope here is obtained by taking statistics beforehand. Specifically, for example, it may be set based on the experimental result shown in FIG. 7 described above. When the experimental result shown in FIG. 7 is followed, for example, the absolute value of the initial value of the parameter a for the L component when the graph is convex is:
It may be set to 0.8 or more. However, if the absolute value of the value is too large at this time, the estimation error of the result of the first estimation process becomes large. Therefore, for example, it is better to appropriately set the absolute value within the range of about 10.0 or less.

FIG. 13 is an explanatory diagram of a specific example of parameter setting. From the experimental results shown in FIG. 7, for example, FIG.
The parameters shown in 3 can be set. Here, in the column of the quantization coefficient SF of FIG. 13, for example, JP
Not the quantized coefficient SF generally called by EG or the like, but the coefficient actually multiplied by the quantized matrix is shown.

The values shown in FIG. 13 are determined on the assumption that the graph is convex upward. The values shown in FIG. 13A show an example of the parameters when the number of coding iterations is allowed only twice. Here, for the tilt parameter a, the maximum absolute value of the values that can be read from the experimental results shown in FIG. 7 was set. In addition, FIG.
The value shown in (B) shows an example of a parameter when the number of repetitions of encoding is permitted three times or more. In this case, a value considered to be an average value was set from the experimental results shown in FIG. 7. For the parameter SF, see FIG.
In both cases (A) and (B), the average value of the quantization coefficient SF that gives a compression rate of 5 was set.

Initial values of parameters as shown in FIG. 13 are set in advance, and at the first time of the estimation process,
The constant presentation unit 7 gives these values to the quantization width calculation unit 9 as parameter data 20. Only one of the parameter values shown in FIGS. 13A and 13B may be set, or the parameter values may be switched depending on the instructed number of repetitions.

Further, FIG. 13 shows the parameters when the graph is convex upward, but when the graph is convex downward, the slope parameter a when the number of iterations shown in FIG. As an example, the minimum absolute value of the values that can be read from the experimental results shown in FIG. 7 may be set.
These parameters can be set according to the shape of the graph, or both the parameters when the graph is convex upward and the parameters when the graph is convex downward can be set. It may be configured to select according to the shape of the graph near the quantized coefficient to be obtained.

By the way, the number of repetitions of encoding can be set to 2 or more, and to any number of times. By increasing the number of times, the code amount becomes closer to the target code amount and the image quality is improved. Moreover, the processing speed can be improved by reducing the number of times. Further, it may be configured to judge whether or not to terminate the iterative process based on the error in the code amount. In this case, there is no idea what the number of iterations will be. For example, there are cases where it only needs to be performed twice, and cases where the number of times is three or more. Therefore, if the estimation process is started by setting the initial values of the parameters similar to the case where the number of iterations is 2, the number of iterations can be dealt with.

Which side the graph is convex can be determined according to the target code amount. As shown in FIG. 5 described above, the overall tendency of the graph is a convex shape when the code amount is large and a convex shape when the code amount is small. Using this property, for example, when the graph is used only with high image quality, it can be initially assumed that the graph is convex. Of course, if there are three or more points of estimation data, the shape of the graph can be determined from those points, and such a configuration may be adopted.

For example, the present invention is based on JPEG Baseli.
When the present invention is applied to a ne-type image encoding apparatus and high-quality encoding is performed with a target compression rate of 5, as a result of experiments on some images, it was found that the quantization rate of any image around the compression rate is 5. It was found that the graph of the logarithm of the conversion factor SF and the logarithm of the code amount CL was convex upward. In such a case, the shape of the graph may be convex upward, and the parameter value may be set or the estimation data may be selected.

Next, the process of estimating the quantized coefficient SF based on the initial value of the estimation parameter and how to take it will be described. FIG. 14 is an explanatory diagram of a process when the final quantized coefficient SF is determined by two estimation processes when the graph is convex upward. First, the initial value SF ₀ of the quantization coefficient and the slope parameter a are set. As the initial value SF ₀ of the quantization coefficient, an average value obtained in advance may be used. For example, it is conceivable to collect SF statistics for various images in advance and use a representative value such as an average value from the statistics. Further, the inclination parameter a is set to a value that is steeper than the assumed inclination as described above. Then, the quantization width is set based on the initial value SF ₀ of the quantization coefficient, the quantization process and the variable length coding process are performed, and the code amount CL ₀ is obtained. Code amount CL obtained in this way
_{From 0} , the initial value SF ₀ of the quantized coefficient and the value of the slope parameter a, the coefficient b of the equation (1) is determined. As a result, the straight line L1 in FIG. 14 is determined. This straight line L1 is
It is a straight line set so as to have a steeper slope than the assumed slope. The quantized coefficient SF _1, which is the target code amount, is obtained using the equation showing this straight line. This quantization coefficient SF ₁
Is the quantized coefficient obtained by the first estimation.
At this time, since the straight line L1 has a steep slope in the first estimation process, the quantization coefficient SF ₁ is a value between the target quantization coefficient SF and the initial value SF ₀ of the quantization coefficient.

In the second estimation process, first, the quantization width is set using the quantization coefficient SF ₁ , and the quantization process and the variable length coding process are performed to obtain the code amount CL ₁ . Then, code amount CL ₀ corresponding to the quantized coefficient SF _0, and obtains the equation indicating the straight line with the code amount CL ₁ and the corresponding quantization coefficient SF _1. In FIG. 14, the straight line obtained by the second estimation process is indicated by the straight line L2. The quantization coefficient SF ₂ is estimated using this straight line. That is, the quantization coefficient SF ₂ when the target code amount is obtained on this straight line is obtained.

The straight line L2 is the quantized coefficients SF ₀ and SF.
_{Since 1} exists on the same side of the quantized coefficient SF to be obtained together, it always exists above the graph in the vicinity of the quantized coefficient SF to be obtained. Therefore, the code amount obtained by performing the quantization process and the variable length coding process using the quantization coefficient SF ₂ obtained in the second estimation process is always equal to or smaller than the target code amount.

The quantized coefficient SF thus obtained
By determining the quantization width using ₂ , and performing the quantization processing and the variable length coding processing, the code amount of the obtained code is close to the target code amount, and the error does not exceed the target code amount.

FIG. 14 illustrates the case where a value larger than the quantized coefficient SF to be obtained is set as the quantized coefficient initial value SF _0.
The same applies when a value smaller than the quantization coefficient SF for which F ₀ is to be obtained is set.

FIG. 15 is an explanatory diagram of a process in the case where the final quantization coefficient SF is determined by a plurality of estimation processes when the graph is convex upward. Although FIG. 14 shows the process of the estimation process twice, a plurality of estimation data can be obtained by repeating this process. Three
If the estimation process is performed more than once, the unevenness of the graph can be determined. The number of repetitions may be set in advance, or may be repeated until the error is within a predetermined value, or these may be combined.

When the graph is convex upward, the combination of estimation data that does not sandwich the quantized coefficient SF to be obtained and gives the result closest to the target code amount is selected. In FIG. 15, two pieces of estimation data having quantization coefficients SF _m and SF _n larger than the quantization coefficient SF to be obtained are selected. A straight line expression is obtained from the two selected estimation data. Here, the straight line L3 in FIG. 15 is obtained. Then, the quantization coefficient SF _p when this straight line has the target code amount may be obtained. At this quantized coefficient SF _p , the straight line L3 exists above the graph. Therefore, the code amount CL _p obtained based on the quantized coefficient SF _p does not exceed the target code amount.

The quantized coefficient SF thus obtained
By determining the quantization width using _p , and performing quantization processing and variable-length coding processing, the code amount of the obtained code is close to the target code amount, and the error does not exceed the target code amount. .

In FIG. 15, although two pieces of estimation data having a quantization coefficient larger than the quantization coefficient SF to be obtained are used, two estimation data having a quantization coefficient smaller than the quantization coefficient SF are used. The same applies when the use data is used.

FIG. 16 is an explanatory diagram of a process in the case where the final quantization coefficient SF is determined by two estimation processes when the graph is convex downward. Similar to the case where the graph is convex upward, first, the initial value SF ₀ of the quantization coefficient and the slope parameter a are set. At this time, the tilt parameter a is
Set the value so that the slope is gentler than the expected slope. Then, the quantization width is set based on the initial value SF ₀ of the quantization coefficient, the quantization process and the variable length coding process are performed, and the code amount CL ₀ is obtained. A straight line L4 in FIG. 16 is determined from the code amount CL ₀ thus obtained, the initial value SF ₀ of the quantization coefficient and the value of the slope parameter a. Using this straight line L4, the quantized coefficient SF ₁ that is the target code amount is obtained. This quantized coefficient SF ₁ is the quantized coefficient obtained by the first estimation process. At this time,
Since the straight line L4 has a gentle slope, the quantized coefficient SF ₁ has a value sandwiching the target quantized coefficient SF together with the initial value SF ₀ of the quantized coefficient. In FIG. 16, since the initial value SF ₀ of the quantized coefficient is smaller than the quantized coefficient SF, the quantized coefficient SF ₁ is larger than the quantized coefficient SF.

In the second estimation process, first, the quantization width is set using the quantization coefficient SF ₁ , and the quantization process and the variable length coding process are performed to obtain the code amount CL ₁ . Then, code amount CL ₀ corresponding to the quantized coefficient SF _0, and obtains the equation indicating the straight line with the code amount CL ₁ and the corresponding quantization coefficient SF _1. In FIG. 16, the straight line L5 represents the straight line obtained in the second estimation process. Using this straight line, the quantization coefficient SF ₂ that is the target code amount is obtained.

The straight line L5 is the quantized coefficients SF ₀ and SF.
_{Since 1} exists at the position sandwiching the quantized coefficient SF to be obtained, when the graph is convex downward, the quantized coefficients SF ₀ and SF
_{Between 1} is always above the graph. for that reason,
The actual code amount obtained by performing the quantization process and the variable length coding process using the quantization coefficient SF ₂ obtained in the second estimation process is always equal to or smaller than the target code amount.

The quantized coefficient SF thus obtained
By determining the quantization width using ₂ , and performing the quantization processing and the variable length coding processing, the code amount of the obtained code is close to the target code amount, and the error does not exceed the target code amount.

FIG. 16 illustrates the case where a value smaller than the quantized coefficient SF to be obtained is set as the quantized coefficient initial value SF _0.
The same applies when a value larger than the quantization coefficient SF for which F ₀ is to be obtained is set.

FIG. 17 is an explanatory diagram of a process in the case where the final quantization coefficient SF is determined by a plurality of estimation processes when the graph is convex downward. Although FIG. 16 shows the process of the estimation process twice, a plurality of estimation data can be obtained by repeating this process. When the graph is convex downward, the quantization coefficient SF to be obtained
Among the combinations of estimation data sandwiching between, the combination closest to the target code amount is selected. In FIG. 17, two pieces of estimation data having a quantized coefficient SF _m larger than the quantized coefficient SF to be obtained and a quantized coefficient SF _n smaller than the quantized coefficient SF to be obtained are selected. Selected 2
A straight line formula is obtained from the three estimation data. here,
The straight line L6 in FIG. 17 is obtained. Then, the quantization coefficient SF _p when this straight line has the target code amount may be obtained. In the case of a downward convex graph, this quantization coefficient SF _p
In, the straight line L6 exists above the graph. Therefore, the code amount CL _p actually obtained based on the quantized coefficient SF _p does not exceed the target code amount.

Quantization coefficient SF obtained in this way
By determining the quantization width using _p , and performing the quantization process and the variable length coding process, the code amount of the obtained code is close to the target code amount, and the error does not exceed the target code amount. Absent.

[0081]

As is apparent from the above description, according to the present invention, in the code amount control of lossy encoding, the code amount is made close to the target code amount and the code amount does not exceed the target code amount. Can be controlled to. This allows
The code amount of the encoded image can be increased as much as possible within the target code amount, and deterioration of the image quality can be suppressed to the minimum.

Also, the logarithm of the quantization coefficient SF and the code amount CL
The shape of the logarithmic graph of can be either upwardly convex or downwardly convex. Further, by changing the initial value of the quantization coefficient SF, it becomes possible to obtain a more accurate quantization width within a desired number of estimation processes. Furthermore, by switching the slope parameter a according to the shape of the graph and using it, the code amount can be brought close to the target code amount within a range that does not exceed the target code amount by only two estimation processes, and the processing efficiency is improved. Further, there is an effect that the image quality can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an image encoding device of the present invention.

FIG. 2 is a block diagram showing an example of an image coding apparatus using a conventional variable length coding method.

FIG. 3 is a flowchart showing an example of the overall operation of an image coding apparatus using a conventional variable length coding method.

FIG. 4 is a flowchart showing an example of an operation of an encoding process in an image encoding device using a conventional variable length encoding system.

FIG. 5 Quantization coefficient SF and code amount CL obtained from experiments
It is a graph which shows an example of the relationship of.

FIG. 6 is an explanatory diagram when a graph of the logarithm of the quantization coefficient SF and the logarithm of the code amount CL is viewed microscopically.

FIG. 7 is a graph showing variations in the value of the inclination a due to differences in images.

FIG. 8 is an explanatory diagram of the principle of the Newton-Raphson method.

FIG. 9 is an explanatory diagram of an estimation error in the case where the graph has a curve that is convex upward.

FIG. 10 is an explanatory diagram of an estimation error when the graph has a downwardly convex curve.

FIG. 11 is a flowchart illustrating an example of an operation in an embodiment of the image encoding device of the present invention.

FIG. 12 is a flowchart showing an example of an operation of tilt setting processing.

FIG. 13 is an explanatory diagram of a specific example of parameter setting.

FIG. 14 is an explanatory diagram of a process in the case where the final quantization coefficient SF is determined by two estimation processes when the graph is convex upward.

FIG. 15 is an explanatory diagram of a process in a case where a final quantization coefficient SF is determined by a plurality of estimation processes when the graph is convex upward.

FIG. 16 is an explanatory diagram of a process in a case where a final quantization coefficient SF is determined by two estimation processes when the graph is convex downward.

FIG. 17 is an explanatory diagram of a process of determining a final quantization coefficient SF by a plurality of estimation processes when the graph is convex downward.

[Explanation of symbols]

1 ... Image conversion unit, 2 ... Quantization unit, 3 ... Variable length coding unit,
4 ... Code amount calculation unit, 5 ... Control unit, 6 ... Parameter selection unit, 7 ... Constant presentation unit, 8 ... Slope calculation unit, 9 ... Quantization width calculation unit, 10 ... Quantization width estimation unit, 11 ... Image data , 12
... Transformed data, 13 ... Quantized data, 14 ... Code data, 15 ... Partial code amount data, 16 ... Total code amount data,
17, 18, 19 ... Control data, 20, 21 ... Parameter data, 22 ... Quantization width data.

 ─────────────────────────────────────────────────── ─── Continuation of front page (72) Setsu Kunitake 2274 Hongo, Ebina City, Kanagawa Prefecture, Fuji Xerox Co., Ltd. ) Inventor Isao Uesawa 2274 Hongo, Ebina City, Kanagawa Prefecture Fuji Xerox Co., Ltd.

Claims (10)

[Claims] 1. An image conversion means for converting an image and outputting converted data, a quantizing means for quantizing in accordance with a quantization width given to the converted data, and outputting quantized data, and the quantizing means. Variable length encoding means for performing variable length encoding on the encoded data, code amount calculating means for accumulating the code amounts created by the variable length encoding means to obtain the code amount of the unit image, and the code amount The control means for comparing the code amount of the unit image created by the calculating means with the preset target code amount and performing control according to the result, and the control result from the control means The quantization coefficient is calculated based on the shape of the graph of the logarithm of the quantized coefficient and the logarithm of the corresponding code amount and the target code amount, and the quantization width is calculated based on the quantized coefficient and the quantization is performed. Quantum given to means An image coding apparatus having a coding width estimating means. 2. When the shape of the graph of the logarithm of the code amount corresponding to the logarithm of the quantization coefficient is convex upward, the quantization width estimation means is larger than the target code amount and closest to the target code amount. Calculate the quantization coefficient and the quantization width by selecting two coding results that give a code amount or two coding results that give a code amount that is less than the target code amount and is closest to the target code amount. The image coding apparatus according to claim 1, wherein 3. The quantization width estimation means is larger than the target code amount and closest to the target code amount when the shape of the graph of the logarithm of the code amount corresponding to the logarithm of the quantization coefficient is downwardly convex. Calculating the quantization coefficient and the quantization width by selecting one encoding result that gives a code amount and one encoding result that gives a code amount that is less than the target code amount and closest to the target code amount. The image coding apparatus according to claim 1, wherein: 4. The quantization width estimation means sets a parameter selection means for adaptively selecting a parameter creation method according to the control from the control means, and a constant predetermined according to the control from the parameter selection means. Constant presenting means that provides as initial values of the slope parameter and the quantization coefficient,
Slope calculation means for calculating the slope parameter based on the shape of the graph of the logarithm of the quantized coefficient and the logarithm of the corresponding code amount according to the control from the parameter selection means, and the constant presenting means or the slope calculation means. The image coding apparatus according to claim 1, further comprising a quantization width calculation unit that calculates the quantization coefficient and the quantization width based on a gradient parameter and the target code amount. 5. The constant presenting means indicates an absolute value of an initial value of the slope parameter given to the quantization width calculating means,
If the graph of the logarithm of the quantization coefficient and the logarithm of the corresponding code amount is convex upward, it is larger than the expected value, and if the graph is convex downward, it is smaller than the expected value. The image encoding device according to claim 4, wherein the image encoding device is set. 6. The image according to claim 4, wherein the constant presenting means sets the initial value of the quantization coefficient given to the quantization width calculating means to an average value obtained in advance. Encoding device. 7. The constant presenting means prepares a plurality of parameters to be given to the quantizing width calculating means, and selectively uses the parameters according to a target code amount. Image encoding device. 8. The control means repeats the estimation process and the encoding process on the unit image until the difference between the code amount of the unit image and the target code amount reaches a certain allowable value, and then outputs the code. The image coding apparatus according to claim 1, wherein the image coding apparatus is controlled as follows. 9. The image conversion means comprises conversion data storage means for storing the conversion data, and the conversion data can be repeatedly read out as necessary, and the image conversion means can read out the conversion data repeatedly. The image encoding device according to 1. 10. The variable-length coding means includes code storage means for storing the created code, and has a code that is the latest code or a code quantity that is closest to the target code quantity, or is the code that is closest to or less than the target code quantity. The image coding apparatus according to any one of claims 1 to 9, further comprising a storage unit that stores a code having a code amount and can read the code when necessary.