RU2607624C1 - Method of controlling output file size from image jpeg compression - Google Patents

Abstract

FIELD: data processing.

SUBSTANCE: invention relates to compressing digital images. Method of controlling speed in JPEG compression of a digital image includes the stages, at which received are coefficients of discrete cosine transform (DCT) obtained as the result of direct DCT application to a digital image; generated are quantization tables (QT) using the received DCT coefficients by iterative estimating the scaling factor to create scaled QT, checking whether the first condition of the iterations completion is true, herewith the first condition of the iterations completion is considered to be true, when the complete actual number of bits calculated for all the colour planes is within the range of tolerance defined, at least, on the basis of the target size of the output file for the JPEG-compressed image, and if the said condition is false, the scaling factor is updated, as well as the said variables, including boundary values of the first range of the scaling factor.

EFFECT: providing a digital image JPEG compression speed control by means of generating a set of quantization tables.

13 cl, 9 dwg

Description

FIELD OF THE INVENTION

The present invention relates generally to the field of digital image compression, and in particular to image compression rate (RC) control according to the standard of the Joint Photographic Expert Group (JPEG).

BACKGROUND OF THE INVENTION

JPEG is a very popular image compression standard, but most of its well-known implementations have a lack of speed control functionality. JPEG speed control is the process of configuring the internal parameters of a JPEG compression process so that JPEG compression of the input image creates a compressed JPEG file with a user-defined size. The need to create a compressed JPEG file with a user-defined size often arises in many areas, for example, when the image captured by the camera is compressed into a JPEG file for storage on a flash memory card with a limited capacity or for transmission over a network channel with limited bandwidth . Most of the known speed control JPEG implementations have the disadvantage of overly high complexity or low speed control accuracy.

The main speed control method provided by the JPEG image compression standard relates to defining quantization tables (QT), which are used by JPEG compression (JPEG compressor) to quantize the discrete cosine transform (DCT) coefficients applied to the color planes of the input image. Since the task of speed control is to configure the internal JPEG compression parameters so that the result of the compression process is a file having a size substantially close to a user-defined value, this task is usually solved by generating and using a set of suitable QTs.

US patent 6192157 ("Modifications of postscript adaptive data compression (ADC) for 3 plane, 8 bit color images, JPEG lossy compression, and variable Q factors"), Hewlett-Packard, describes a JPEG image compression method adapted to create output files, the size of which corresponds to the output buffer of a predetermined size. This method comprises a variable Q factor, which is a scaling factor for generating a scaled version of the standard QT JPEG, and the JPEG compressor is configured to increase the Q coefficient to a level sufficient to ensure that the size of the output JPEG file matches the output buffer. Although this method provides fairly accurate control of the size of the output file, it has the disadvantages of high computational complexity due to several repeated compressions necessary to find a suitable Q coefficient, and the absence of improvements in the quality of the reconstructed image, possible through the use of individually generated QTs.

US patent 7567722 ("Dynamically scaled file encoding"), Qualcomm Inc., describes a JPEG image compression method comprising a speed control method based on a two-pass process for determining the standard QT scaling factor, which provides JPEG compression to a predetermined file size. The fraction of nonzero quantized DCT coefficients is used to adjust the initial scaling factor. Although this method provides low computational complexity and acceptable accuracy of speed control, it does not provide improved image quality compared to JPEG compression, which uses scaled versions of standard QT JPEG. In addition, its computational complexity and speed control accuracy can be further improved by using decimated image compression and a two-parameter prediction model instead of a single-parameter prediction model.

US Pat. No. 7,574,066 ("Method for image compression, related system and computer product therefor"), STMicroelectronics Ltd., describes a JPEG image compression method comprising a speed control method based on iteratively adjusting the scaling factor of standard JPEG QTs to achieve a predetermined output file size. The disadvantages of this method are the high computational complexity and insufficient quality of the reconstructed image.

US patent 5724453 ("Image compression system and method having optimized quantization tables"), Wisconsin Alumni Research Foundation, describes a JPEG image compression method containing a speed-distortion optimization (RDO) method based on a dynamic programming technique (DP), this method provides both precise speed control and improved image quality by generating RD-optimal QTs. Although this method has good performance in terms of both computational complexity and the quality of the reconstructed image, it is characterized by a lack of effective adaptability to the image, since it uses the average entropy of the quantized DCT coefficients to estimate the output bit rate (bit rate), which is not accurate in general case.

US patent 7567722 and US patent 5724453 can be considered as the closest analogues for this application.

SUMMARY OF THE INVENTION

An object of the present invention is to provide a method for configuring internal parameters (in particular, quantization tables) of a JPEG compression process so that JPEG compression of the input image results in a compressed JPEG file having a user-defined size.

According to a first aspect of the present invention, there is provided a speed control (RC) method within JPEG compression of a digital image. According to the RC method, discrete cosine transform (DCT) coefficients obtained from applying direct DCT to a digital image are received. Then, quantization tables (QTs) are generated using the received DCT coefficients by iteratively evaluating the scaling factor to create scaled QTs, checking if the first iteration termination condition is true, and the variables associated with the coefficient are involved in the first iteration termination condition scaling, and if said condition is false, then the scaling factor and the said variables are updated.

The RC method may further comprise, before the QT generation step, the steps of: collecting DCT histograms, the DCT histogram collecting step comprising, for each of the color planes: counting the number of occurrences of each DCT coefficient value for each DCT subband; and initializing the target size of the output file for the image compressed according to JPEG, the scaling factor, the boundary values of the range of the reactive number of bits (first number of bits), the lower and upper tolerances on the reactive number of bits, the boundary values of the range of the reactive scaling factor (first range of the scaling factor), corresponding to the boundary values of the range of the reactive number of bits, and the accuracy parameter of the scaling factor.

The step of evaluating the scaling factor may include the steps in which: initialize the boundary values of the range of the predictive number of bits (second number of bits), the lower and upper tolerances for the predictive number of bits and the boundary values of the range of the predictive number of bits (second range of the scaling factor) corresponding to the boundary values of the range predictive number of bits; iteratively scale the predetermined base QTs based on the scaling factor to obtain scaled QTs, calculate the total predictive number of bits as the sum of the predictive number of bits calculated individually for each of the color planes, the predictive number of bits resulting from applying scaled QT to JPEG compression , obtained by using the parameters of the prediction models, check whether the second condition for completing the iterations is true, the second condition The completion of the iterations is true when the total predictive number of bits falls within the tolerance range for the predictive number of bits or the width of the range of the predictive scaling factor becomes less than the accuracy parameter of the scaling factor, and the tolerance range for the predictive number of bits is determined based on the target size of the output file and the lower and the upper tolerances for the predictive number of bits, and if the second condition for completing the iterations is false, then update yayut boundary predictive value quantity range bits, and boundary values of scaling the scaling factor predictive coefficient based on the full range of the predictive number of bits using the method of false position for finding the root.

Predefined base QTs are preferably standard JPEG QTs. The predictive number of bits for a particular color plane is preferably calculated using a polynomial prediction model based on prediction model parameters and prediction parameters for that color plane, the prediction parameters being calculated using quantized DCT coefficients obtained by quantizing DCT coefficients using scaled QTs.

The RC method may further comprise, after the step of estimating the scaling factor, quantizing and counting the bits by using scaled QTs in standard JPEG quantization and entropy coding processes, modified so that only the bits output during entropy coding are counted without outputting compressed data. The bit counting is performed separately for each of the color planes, and as a result, a set of actual number of bits of the color planes is obtained, and the bit count further comprises: calculating the total actual number of bits by summing the actual number of bits of the color planes for all color planes. The first condition for completing iterations is true when the total actual number of bits falls within the tolerance range for the reactive number of bits, or the range width of the reactive scaling factor becomes less than the accuracy parameter of the scaling factor, or the number of iterations reaches a predetermined maximum number. The tolerance range for the reactive bit count is determined based on the target size of the output file and the lower and upper tolerances for the reactive bit count. At the stage of updating the scaling factor and the associated variables, the boundary values of the reactive number of bits range, the boundary values of the reactive scale factor range and the scale factor are updated based on the total actual number of bits using the false position method to find the root. Moreover, the RC method further comprises, after the step of updating the scaling factor and the corresponding variables, the step of: updating the set of prediction model parameters for each color plane using the set of actual number of bits of the color planes, scaled QT and DCT histograms; and, if each set of parameters of the prediction model is updated, then the predetermined base QTs are scaled based on the scaling factor and return to the step of quantizing and counting bits; otherwise, they return to the step of estimating the scaling factor.

The step of updating the parameters of the prediction model preferably comprises the steps of: calculating a prediction parameter for each of the color planes using scaled QT and DCT histograms; adaption batteries are updated using the prediction parameter and the normalized actual number of bits of the color plane for a given color plane, the preliminary adaptation batteries being initialized to zero;

if the system of linear equations formed from the said adaptation batteries is joint, then the parameters of the prediction model are calculated by solving the said system of linear equations, otherwise, the parameters of the prediction model are not updated.

The RC method may further comprise stages in which: one of the RC operation modes is selected, the RC operation modes include: a fast mode providing the best performance; balanced mode that provides a compromise between performance and image quality; and a quality mode that provides the best image quality; calculating the total distortion associated with the scaled QT for each of the color planes; calculate the target number of bits of the color planes for each of the color planes using the set of actual number of bits of the color planes.

If the balanced mode is selected, the RC method further comprises: performing speed-distortion optimization of the balanced mode to generate QT optimized for speed-distortion (RDO QT), wherein the speed-distortion optimization of the balanced mode comprises, for each of the color planes, the steps on which: initialize the RDO QT records; calculating the records of the velocity-distortion (RD) tables using DCT histograms and prediction model parameters, said records being dependent on a variable quantizer; determining a set of locally optimal quantizers using RDO QT entries, RD table entries, and a target number of bits of a color plane calculated for said color plane; searching for a locally optimal DCT subband using a set of locally optimal quantizers to determine the index i * of the locally optimal DCT subband; if a locally optimal DCT subband exists, then the quantizer is updated by setting the i * -th RDO QT record equal to the i * -th locally optimal quantizer, and return to the step of determining the set of locally optimal quantizers.

If a quality mode is selected, then the RC method further comprises: performing speed-distortion optimization of the quality mode for generating an RDO QT, wherein optimizing for speed-distortion of the quality mode comprises, for each of the color planes, the steps in which: the entries of the RD tables are calculated using DCT histograms and parameters of the prediction model, said records being dependent on a variable quantizer; iteratively, calculating the entries of the optimal quantizer-velocity-distortion (RDQ) tables using the RD tables; initialize iterations of obtaining a quantizer based on the records of the RDQ tables and the target number of bits of the color plane calculated for said color plane; iterate the quantizer using the RDQ table entries to create the RDO QT.

The RC method may further comprise the steps of: calculating the total distortion associated with the RDO QT; perform bit counting by using RDO QT in standard quantization and entropy coding processes according to JPEG, modified so that only the bits output by entropy coding are counted without outputting compressed data, and bit counting is performed separately for each of the color planes, and as a result it receives the actual number of bits, and the bit count further comprises the steps of: calculating the final actual number of bits by summing the fact number of bits for all color planes, if the total distortion associated with the QT RDO is less than the total distortion associated with the scaled QT and the final total actual number of bits falls within the tolerance range for the final number of bits, with the final tolerance range for the final the number of bits is determined based on the target size of the output file and predefined tolerances on the final number of bits, then output the RDO QT as the generated QT, otherwise, output the scale sorbed QT as QT generated.

According to a second aspect of the present invention, there is provided a computer-readable storage medium on which computer-executable instructions are stored. Computer-executable instructions stored on a computer-readable storage medium, when executed by a processor, instruct the processor to execute the RC method according to the first aspect of the present invention.

According to a third aspect of the present invention, there is provided a device for performing RC in JPEG compression of a digital image. This device contains: one or more processors; and a memory in which machine-executable instructions are stored. Computer-executable instructions stored in memory, when executed by one or more processors, instruct one or more processors to execute the RC method according to the first aspect of the present invention.

According to a fourth aspect of the present invention, there is provided a JPEG method for compressing a digital image. Said method comprises the steps of: performing color space conversion and direct DCT; performing a method according to a first aspect of the present invention; perform quantization; perform entropy coding; and output the compressed JPEG file.

Unlike prior art methods, the present invention provides an optimal combination of low computational complexity, good speed control accuracy and, when using quality improvement modes, a marked improvement in the quality of the reconstructed image.

Brief Description of the Drawings

FIG. 1 - a sequence of JPEG compression operations with RC capability;

FIG. 2 is a flowchart of a speed control process;

FIG. 3 is a flowchart of adaptation of the predictor of the number of bits;

FIG. 4 is a flowchart for estimating a predictive scaling factor;

FIG. 5 is a flowchart of calculating a predictive number of bits;

FIG. 6 is a flowchart of updating a prediction model;

FIG. 7 is a flowchart of a balanced mode RDO for a single color plane;

FIG. 8 is a flowchart of operation of an RDO quality mode for one color plane;

FIG. 9 is a block diagram illustrating an apparatus according to an embodiment of the present invention.

Detailed Description of Embodiments

A method for controlling the size of the output image compression JPEG file is shown in FIG. 1. An input digital image created, for example, by means of a digital camera or a document scanning device, is subjected to color space conversion (for example, RGB-B-YCbCr conversion, if necessary) and direct DCT (101). The resulting DCT coefficients are processed by a speed control process (102) that outputs a set of QT (one QT per color plane). The speed control process is a new method that meets the essence of the present invention. The goal of the speed control process is to generate a QT set so that JPEG compression, where this QT set is to be used, produces an output file compressed according to JPEG, having a file size substantially close to a user-defined value. The output QT set is used in quantization (103) to create quantized DCT coefficients, for which entropy encoding is then performed (104) to create an output compressed JPEG file for, for example, storing it on a flash card with limited memory capacity or transmitting via a network channel limited bandwidth.

The speed control process (102) is shown in FIG. 2. First, when adapting (201) the bit number predictor, the bit number predictor is adapted using the DCT coefficients and a set of scaled QTs and a set of prediction model parameters are output. Then, the RC operation mode is switched: if the user selects the fast RC operation mode, the set of scaled QTs created in step 201 are output. Since the set of scaled QTs are generated in step 201 so that JPEG compression using this set of QT creates an output a compressed JPEG file having a file size substantially close to a user-defined value, a speed control goal is achieved by outputting said QT set. Otherwise, if the user selects one of two RCO RDO modes of operation, then the total distortion associated with the scaled QT (202) is calculated, and the calculated total distortion is then compared with the total distortion associated with the QT RDO. Then, bit quota distribution among the color planes (203) is performed to determine the target number of bits for each of the color planes, which is introduced into the subsequent RDO process. After that, the RDO mode is switched: depending on the user choice, either the balanced mode RDO (204) or the quality mode RDO (205) are performed. The balanced mode RDO and quality mode RDO processes both use the set of prediction model parameters from step 201. The output of the RDO process of any mode is the set of RDO QTs that are used in bit counting and total distortion calculation using RDO QT (206). The final actual number of bits obtained and the total distortion associated with the RDO QT obtained in step 206 are compared, respectively, with the tolerance range for the final number of bits and the total distortion associated with the scaled QT (obtained in step 202). If the final actual number of bits associated with the RDO QT falls within the tolerance range for the final number of bits and the total distortion associated with the RDO QT is less than the total distortion associated with the scaled QT, then the RDO QT set is output. Otherwise, the set of scaled QTs created in step 201 are output. Since both processes of the two RDO modes are designed to generate the QT set so that JPEG compression using this QT set creates an output JPEG compressed file having a file size substantially close to a user-defined value, the speed control goal is achieved by outputting the QTs mentioned, while comparing the total distortion after step 20–6 results in a QT set that provides the best quality for subsequent JPEG compression tiya.

A detailed description of the speed control process (102) is given below.

The adaptation (201) of the bit predictor is used to process the DCT coefficients and output a set of scaled QTs and a set of prediction model parameters. Adaptation of the predictor of the number of bits is critical to achieve a combination of low computational complexity and good speed control accuracy. A scaled QT is preferably a scaled version of some basic QT, for example, one of the standard JPEG QTs given in Appendix K of the ISO / IEC JTC1 10918-1 JPEG specification:

является значением квантователя для i-ой субполосы DCT в базовой QT, и clip (х, x_min, x_max) определена следующим образом:where i is the DCT subband index, S is the scaling factor, Q _min = 1, Q _max = 255 for basic JPEG compression, or 65535 for advanced JPEG compression,

is the quantizer value for the i-th DCT subband in the base QT, and clip (x, x _min , x _max ) is defined as follows:

:The adaptation (201) of the bit predictor is shown in FIG. 3. When collecting (301) DCT histograms, the number of occurrences of each DCT coefficient value for each DCT subband for each color plane is counted to collect DCT histograms

:

является значением коэффициента i-ой субполосы DCT j-ого блока DCT 8×8 k-ой цветовой плоскости,

является количеством блоков DCT 8×8 в k-ой цветовой плоскости, N_c является количеством цветовых плоскостей во входном изображении, q_s является конфигурируемым параметром точности коэффициента DCT, s является значением коэффициента DCT, дискретизованным с точностью

, T_max является максимальным значением коэффициента DCT, зависящим от формата данных входного изображения (например, T_max=2¹⁰ для 8-битных данных элемента изображения входного изображения), и [х] определяется как целое значение, ближайшее к х. Затем, гистограммы DCT используются по всему процессу управления скоростью.where i is the DCT subband index, k is the color plane index, j is the 8 × 8 DCT block index,

is the coefficient value of the i-th DCT subband of the j-th DCT block of 8 × 8 k-th color plane,

is the number of 8 × 8 DCT blocks in the kth color plane, N _c is the number of color planes in the input image, q _s is a configurable accuracy parameter of the DCT coefficient, s is a DCT coefficient value sampled with accuracy

, T _max is the maximum value of the DCT coefficient, depending on the input image data format (for example, T _max = 2 ¹⁰ for 8-bit image element data of the input image), and [x] is defined as the integer value closest to x. Then, DCT histograms are used throughout the speed control process.

Upon initialization (302) of the reactive context, the initial values of the variables and constants of the reactive context are set:

- B _t is the total target number of bits specified as the user-defined required JPEG file size in bits, minus the number B _{m of} JPEG marker bits;

- B _RL and B _RU are, respectively, the lower and upper limit of the range of the reactive number of bits (which are initialized, respectively, by the minimum number of bits B _min and the maximum number of bits B _max );

- ε _RL and ε _RU are, respectively, the lower tolerance on the reactive number of bits and the upper tolerance on the reactive number of bits (for example, initialization of 0.01 and 0.01 corresponds to an accuracy of the reactive number of bits equal to 1%);

- S is the current scaling factor used to scale the base QT (which is preferably initialized to 100);

- S _RL and S _RU are, respectively, the lower boundary and the upper limit of the range of the reactive scaling factor corresponding to B _RL and B _RU ;

- ΔS is a parameter of the accuracy of the scaling factor.

The reactive context is then used throughout the adaptation of the predictor of the number of bits.

When evaluating (303) the predictive scaling factor, the value of the scaling factor is estimated using the method shown in FIG. 4. When initializing (401) the predictive context, the initial values of the variables and constants of the predictive context are set:

- B _PL and B _PU are, respectively, the lower and upper boundary of the range of the predictive number of bits (which are initialized in B _RL and B _RU );

- ε _PL and ε _PU are, respectively, the lower tolerance for the predictive number of bits and the upper tolerance for the predictive number of bits (for example, their initialization of 0.01 and 0.01 corresponds to the accuracy of the predictive number of bits equal to 1%);

- S _PL and S _PU are, respectively, the lower bound and the upper bound of the predictive scaling factor range corresponding to B _PL and B _PU (which are preferably initialized to S _RL and S _RU , respectively);

- B _t , S and ΔS are shared by both reactive and predictive contexts, so their values do not change.

The predictive context is then used throughout the assessment of the predictive scaling factor.

When scaling base QTs (402), a set of scaled QTs is created according to equation (1) using the current scaling factor S. By calculating the predictive number of bits (403), the number of bits obtained using the scaled QT set from step 402 in JPEG compression is used, where the current parameters of prediction models. The prediction process is shown in FIG. 5.

. Параметр ρ^(k) определяется как доля квантованных коэффициентов DCT с нулевыми значениями и вычисляется для всех цветовых плоскостей с использованием следующей формулы:The calculation of the prediction parameters (501) gives the prediction parameters for all color planes. The prediction parameter for the kth color plane can be either 1-ρ ^(k) or the entropy averaged over the coefficients

. The parameter ρ ^(k) is defined as the fraction of quantized DCT coefficients with zero values and is calculated for all color planes using the following formula:

является значением квантователя i-ой субполосы DCT для k-ой цветовой плоскости.Where

is the quantizer value of the i-th DCT subband for the k-th color plane.

определяется как среднее значение всех энтропий субполосы DCT квантованных коэффициентов DCT и вычисляется для всех цветовых плоскостей с использованием следующей формулы:Coefficient Averaged Entropy

defined as the average of all the entropies of the DCT subband of the quantized DCT coefficients and calculated for all color planes using the following formula:

where z is the value of the quantized DCT coefficient.

для всех цветовых плоскостей с использованием полиномиальной модели предсказания:The calculation of the outputs of the prediction models (502) uses the prediction parameters to calculate the predictive number of bits

for all color planes using a polynomial prediction model:

вычисленным с использованием уравнения (4) или (5), соответственно), m является индексом параметра модели предсказания,

является m-ым параметром модели предсказания для k-ой цветовой плоскости, и N_m является количеством параметров модели предсказания. Суммирование всех выходов моделей предсказания (503) создает полное предиктивное количество битов В_р для всех цветовых плоскостей:where P ^(k) is the prediction parameter for the kth color plane (which is either 1-ρ ^(k) , or,

calculated using equation (4) or (5), respectively), m is the index of the prediction model parameter,

is the m-th parameter of the prediction model for the k-th color plane, and N _m is the number of parameters of the prediction model. The summation of all the outputs of the prediction models (503) creates the total predictive number of bits B _p for all color planes:

, полученный масштабированием базовых QT с использованием текущего коэффициента масштабирования S. Иначе, при обновлении (404) предиктивного коэффициента масштабирования используют полное предиктивное количество битов В_р для обновления переменных предиктивного контекста S_PL, S_PU, B_PL, и B_PU и текущего коэффициента масштабирования S с использованием метода ложного положения для нахождения корня посредством следующей формулы:If the total predictive number of bits falls within the tolerance range for the predictive number of bits, namely, B _t (1-ε _PL ) ≤B _p ≤Bt (1 + ε _PU ), or the width of the range of the predictive scaling factor becomes less than the scaling factor accuracy parameter namely, | S _PU -S _PL | <ΔS, then a set of scaled QTs is issued

obtained by scaling the base QTs using the current scaling factor S. Otherwise, when updating (404) the predictive scaling factor, the full predictive number of bits B _{p is used} to update the predictive context variables S _PL , S _PU , B _PL , and B _PU and the current scaling factor S using the false position method to find the root by the following formula:

After step 404, the process is repeated from step 402 until the above inference condition is satisfied.

; затем, суммирование всех фактических количеств битов цветовых плоскостей дает полное фактическое количество битов В_а:After step 303, quantization and bit counting (304) is performed using standard JPEG quantization and entropy coding processes, modified so that only counted bits are output during entropy coding (without outputting compressed data). The bit counting is performed separately for each of the color planes, and as a result of it, a set of actual numbers of bits of the color planes is formed

; then, the summation of all the actual number of bits of the color planes gives the total actual number of bits In _a :

, полученный масштабированием базовых QT с использованием текущего коэффициента масштабирования S, и набор параметров модели предсказания

. Иначе, при обновлении (305) реактивного коэффициента масштабирования используют полное фактическое количество битов В_а для обновления переменных реактивного контекста S_RL, S_RU, B_RL и B_RU и текущего коэффициента масштабирования S с использованием метода ложного положения для нахождения корня посредством следующей формулы:If the total actual number of bits B _a from step 304 falls within the tolerance range for the reactive number of bits, namely, Bt (1-ε _RL ) ≤B _a ≤B _t (1 + ε _RU ), or the bandwidth of the reactive scaling factor becomes less than the accuracy parameter of the scaling factor, namely, | S _RU -S _RL | <ΔS, or the number of iterations N _iter for quantizing and counting bits (304) reaches a predetermined maximum number of iterations N _max , then a set of scaled QT

obtained by scaling basic QTs using the current scaling factor S and a set of prediction model parameters

. Otherwise, when updating (305) the reactive scaling factor using the full actual number of bits B _and to update the variable jet context S _RL, S _RU, B _RL and B _RU and the current scale factor S using the method of false position for finding the root by the following formula:

и набор гистограмм DCT

для вычисления параметра предсказания Р^(k) (который является либо 1-ρ^(k), либо

, вычисленным с использованием уравнения (4) или (5), соответственно).When updating (306) the prediction models, a set of actual numbers of color plane bits obtained in step 304 is used to update the parameters of the prediction model for each color plane using the method shown in FIG. 6 for one color plane. When calculating (601) the prediction parameters, the current scaled QT is used.

and DCT histogram set

to calculate the prediction parameter P ^(k) (which is either 1-ρ ^(k) , or

calculated using equation (4) or (5), respectively).

(где

является количеством пикселей в k-ой цветовой плоскости) для обновления адаптационных аккумуляторов

и

с использованием следующей формулы:When updating (602) adaptive batteries, the prediction parameter and the normalized actual number of bits of the color plane are used

(Where

is the number of pixels in the k-th color plane) for updating adaptive batteries

and

using the following formula:

As can be seen from the above, adaptation batteries, in fact, are sets of variables (see relation (11)), which variables accumulate (i.e. contain the sum) of the values of other parameters and are used for adaptation.

All adaptive accumulators are initialized to zero at the beginning of step 201. Adaptive accumulators are used to construct an adaptive system of linear equations (SLE):

. Если адаптационная SLE, определяемая уравнением (12), является несовместной, т.е. определитель матрицы

, определяемый как

, равен нулю, то тогда этап обновления модели предсказания для k-ой цветовой плоскости завершают без обновления параметров модели предсказания. Иначе, вычисление параметров модели предсказания (603) выполняют посредством решения адаптационной SLE в отношении набора параметров модели предсказания

, например, посредством метода на основе разложения Холецкого.where m is the index of the equation and l is the index of the unknown variable

. If the adaptive SLE defined by equation (12) is incompatible, i.e. matrix determinant

defined as

is equal to zero, then the stage of updating the prediction model for the k-th color plane is completed without updating the parameters of the prediction model. Otherwise, the calculation of the parameters of the prediction model (603) is performed by solving the adaptive SLE with respect to the set of parameters of the prediction model

, for example, through the method based on Cholesky decomposition.

для k=1, …, N_c обновлен на этапе 306, то тогда процесс повторяют, начиная с этапа 303. Иначе, масштабированием базовых QT (307) создается набор масштабированных QT согласно уравнению (1) с использованием текущего коэффициента масштабирования S. После этапа 307 процесс повторяют, начиная с этапа 304. В другом варианте осуществления, переход к этапу 303 после этапа 306 выполняют, если по меньшей мере одна из моделей предсказания изменена.After step 306, if all the prediction models are changed, namely, each set of parameters of the prediction model

for k = 1, ..., N _{c is} updated in step 306, then the process is repeated starting from step 303. Otherwise, scaling the base QT (307) creates a set of scaled QT according to equation (1) using the current scaling factor S. After step 307, the process is repeated starting from step 304. In another embodiment, the transition to step 303 after step 306 is performed if at least one of the prediction models is changed.

и

, используемые по всей адаптации предиктора количества битов (этап 201), умножают на ƒ_MCU.In another embodiment, to improve the computational performance of adaptation (201) of the bit predictor, collecting (301) DCT histograms and quantizing and counting bits (304) can be performed not for the entire color plane, but for the fraction ƒ _{MCU of} all minimum encoded blocks ( MCU), randomly selected using uniform distribution over the color plane region. In this case, the values of Bt,

and

used throughout the adaptation of the predictor of the number of bits (step 201) is multiplied by ƒ _MCU .

After adaptation (201) of the predictor of the number of bits, the following three RC operation modes are set: fast mode, balanced mode, and quality mode. The RC-capable JPEG compressor user is responsible for selecting the appropriate RC mode of operation.

The fast RC operation mode provides a set of scaled QTs obtained in step 201.

If you select either a quality mode or a balanced RC mode of operation, then this method proceeds to the next steps.

The calculation of the total distortion associated with the scaled QT (202) is performed using the following formula:

, получаемых посредством квантования и подсчета битов (304) для последней итерации адаптации (201) предиктора количества битов, чтобы вычислить целевое количество битов цветовой плоскости

для k-ой цветовой плоскости с использованием следующей формулы:When distributing the quota of bits among the color planes (203), a set of actual numbers of bits of the color plane is used

obtained by quantization and bit counting (304) for the last iteration of adaptation (201) of the number of bits predictor to calculate the target number of bits of the color plane

for the k-th color plane using the following formula:

является долей количества битов для цветоразностных плоскостей.Where

is the fraction of the number of bits for color difference planes.

After step 203, one of two RDO operations is started depending on the choice of the RC operating mode: RDO (204) balanced mode or RDO (205) quality mode.

If a balanced RC mode of operation is selected, then a balanced mode RDO (204) is performed separately for each color plane. The balanced mode RDO process for one color plane is shown in FIG. 7.

с использованием следующей формулы:The number N _{m of} parameters of the prediction model should be set equal to 2 for a balanced RC mode of operation. Initializing (701) RDO QT Sets the Initial Values of RDO QT Records

using the following formula:

и параметры модели предсказания

для вычисления

и

таблиц скорости-искажения (RD) с использованием следующей формулы:where S _RDO is the initial RDO scaling factor. When filling out RD tables (702), DCT histograms are used.

and prediction model parameters

to calculate

and

velocity distortion (RD) tables using the following formula:

является параметром предсказания для i-ой субполосы DCT для k-ой цветовой плоскости (который является либо 1-ρ^(k), либо

, вычисленным с использованием уравнения (4) или (5), где

заменяется на q),

и

являются минимальным и максимальным значениями квантователя для сбалансированного режима функционирования RC, и max(х, у) определяется как максимальное значение среди х и у.where q is a variable quantizer,

is the prediction parameter for the i-th DCT subband for the k-th color plane (which is either 1-ρ ^(k) or

calculated using equation (4) or (5), where

replaced by q),

and

are the minimum and maximum values of the quantizer for a balanced RC mode of operation, and max (x, y) is defined as the maximum value among x and y.

с использованием следующей формулы:A search for locally optimal quantizers (703) defines a set of locally optimal quantizers

using the following formula:

в уравнении (17).It should be noted that the set of locally optimal quantizers may be empty if the constraint cannot be satisfied

in equation (17).

A search for the locally optimal DCT subband (704) determines the index i * of the locally optimal DCT subband using the output from step 703, based on the following formula:

. Иначе, выполняют обновление (705) квантователя посредством установки

равным

.If there is no locally optimal DCT subband, then the current RDO QT is output

. Otherwise, the quantizer is updated (705) by setting

equal

.

After step 705, the process is repeated starting from step 703 until the above withdrawal condition is met.

If the RC performance mode is selected, then the RDO (205) of the quality mode is performed separately for each color plane. The quality mode RDO process for one color plane is shown in FIG. 8.

и

, вместо

и

.The number N _{m of} parameters of the prediction model should be set equal to 2 for the RC performance mode. Filling in the tables RD (801) is performed in the same way as in step 702, but using other minimum and maximum values of the quantizer

and

instead

and

.

,

и

из таблиц оптимального квантователя-скорости-искажения (RDQ), причем в каждой итерации используют следующую формулу:When populating Optimal RDQ (802) tables, RD tables are used to iteratively compute records

,

and

from the tables of the optimal quantizer-velocity-distortion (RDQ), and in each iteration use the following formula:

, и

определяется как наибольшее целое значение, меньшее х.where b is the index of the sampling step of the number of bits, N _R is the number of sampling steps of the number of bits, ΔB is the width of the sampling step of the number of bits, defined as

, and

defined as the largest integer value less than x.

Iterations begin with i = 0 and proceed to calculating the entries of the optimal RDQ tables for index i using the record values for index i-1 based on equation (19).

и накопленного количества битов

с использованием следующей формулы:The initialization of obtaining a quantizer (803) initializes the iteration of obtaining a quantizer by setting the quantizer index i to 63 and setting the index

and accumulated number of bits

using the following formula:

The iteration (804) of obtaining a quantizer is continued using the following iteration step:

. Иначе, процесс повторяют, начиная с этапа 804, пока не будет выполнено приведенное выше условие вывода.If, after step 804, the RDO QT is completely obtained, namely, the quantizer index i reaches -1, then the obtained RDO QT

. Otherwise, the process is repeated from step 804 until the above inference condition is satisfied.

After the RDO operation (204 or 205), bit counting and total distortion calculation are performed using the RDO QT (206). Bit counting is performed in the same manner as in step 304, but using RDO QT instead of scaled QT. The total distortion associated with the RDO QT is calculated using the following formula:

If the total (final) actual number of bits B _a from step 206 falls within the tolerance range for the final number of bits, namely, Bt (1-ε _FL ) ≤B _a ≤Bt (1 + ε _FU ), and the total distortion associated with RDO QT, is less than the total distortion associated with scaled QTs, namely D _RDO <D _scaled , then the RDO QT set is output

. Иначе, выводят набор масштабированных QT

. Допуски по окончательному количеству битов ε_FL и ε_FU выбираются согласно пользовательским требованиям для точности способа RC: например, установка их равными 0,1 и 0,1 соответствует точности окончательного количества битов, равной 10%.

. Otherwise, a set of scaled QTs is derived.

. Tolerances for the final number of bits ε _FL and ε _FU are selected according to user requirements for the accuracy of the RC method: for example, setting them to 0.1 and 0.1 corresponds to an accuracy of the final number of bits of 10%.

In an embodiment where improved computational performance of the speed control process (102) is needed, all operations including processing color-difference color planes (Cb and Cr) can be performed using a combined method. This is accomplished by setting N _c to 2 (rather than 3) for a color image and processing the Cb and Cr data together when collecting (301) DCT histograms and quantizing and counting bits (304). The output of the operation of the speed control process (102) for the combined color difference color plane is QT, which is used to quantize both the color plane Cb and the color plane Cr during quantization (103).

According to a preferred embodiment of the present invention, the following values are set for the above process parameters:

- a set of basic QTs is a set of standard QT JPEGs specified in Appendix K to the specification of the JPEG standard

- q _s = 1

- T _max = 2 ¹⁰ for 8-bit pixel data of the input image

- B _m = 5000 for JPEG compression, where optimization using the Huffman table is not used

- B _m = 3000 for JPEG compression, where optimization using the Huffman table is used

- B _min = 0, B _max = the maximum floating point value supported by the system executing the method

- ε _FL = 0.1, ε _FU = 0.1 for the accuracy of the final number of bits equal to 10%

- ε _RL = ε _FL , ε _RU = ε _FU for the fast operating mode RC

- ε _RL = 0.01, ε _RU = 0.01 for a balanced RC mode of operation and RC quality mode of operation

- S _RL = 2600, S _RU = 0

- ΔS = 1

- ε _PL = 0.01, ε _PU = 0.01

- the prediction parameter is 1-ρ ^(k)

- N _m = 2 (linear prediction model)

- initial values of the parameters of the prediction model:

- for the color plane Y

- for the color plane Cb

- for the color plane Cr

- N _max = 4 for fast operation RC

- N _max = 8 for balanced RC mode of operation and RC quality mode of operation

- S _RDO = 2

=1,

=255-

= 1,

= 255

=1,

=255-

= 1,

= 255

- N _R = 1000

- ƒ _MCU = 0.5 for fast RC operation

- ƒ _MCU = 1 for balanced RC mode and RC quality mode

- a combined method for color difference color planes.

The described method can be implemented as a component of any JPEG compressor that complies with the standard JPEG base profile standard to provide both control of the JPEG compression rate and improvement of the quality of the reconstructed image. Hardware platforms suitable for implementing the described method include, but are not limited to: a personal computer, a cloud server, a mobile phone, a digital camera, a document scanning device, a digital copier, a multifunction printer, etc.

An embodiment of an apparatus configured to implement RC technology according to the present invention is shown with reference to FIG. 9.

The device 900 includes at least a data processing unit 910 and a memory 920 communicatively coupled to the data processing unit 910.

The data processing unit 910 may comprise one or more known processors (CPUs), controllers, ASICs, etc. Memory 920 may comprise various known volatile computer-readable storage devices and media (e.g., various types of RAM) and non-volatile computer-readable storage devices and media (e.g., various types of ROM, optical media (CD, DVD), HDD, flash drives, etc. .).

RC technology, described in detail above, is preferably implemented in the form of software, firmware, which is recorded as machine-executable instructions in computer-readable storage devices and media forming memory 920. This software can be designed and developed using any suitable technology and infrastructure.

The computer-executable instructions, when read by the data processing unit 910 from the memory 920 and executed by the data processing unit 910, instruct the device 900 to execute the RC method according to the present invention described with reference to FIG. 1-8.

The device 900 may be contained in a personal computer, a cloud server, a mobile phone, a digital camera, a document scanner, a digital copier, a multifunction printer, etc.

In an illustrative embodiment of the inventive RC method in a mobile phone comprising the device described above with reference to FIG. 10, a digital image is first captured using a mobile phone’s camera, then it is supplied to the device together with a user-defined file size and RC operation mode and processed to generate a QT set, which is then used in JPEG image compression to create a JPEG compressed file having a certain user size. A JPEG compressed file can, for example, be uploaded later to a social network, where strict restrictions are placed on the size of a JPEG file.

One possible application is the generation of JPEG compressed images suitable for sharing on social networks such as Facebook, Google+, MySpace, Twitter, etc., which impose strict restrictions on the image file size. Another possible application is the generation of a video stream compressed according to the Motion JPEG standard, suitable for transmission over a network channel having limited bandwidth.

It should be understood that the shown embodiments are only preferred, but not the only possible examples of the present invention. On the contrary, the scope of the invention is defined by the following claims and their equivalents.

Claims (67)

1. The method of speed control (RC), performed as part of JPEG compression of a digital image, the method comprising the steps of: receiving discrete cosine transform (DCT) coefficients obtained by applying direct DCT to a digital image; and quantization tables (QT) are generated using the received DCT coefficients by means of iteratively evaluate the scaling factor to create scaled QTs, check whether the first condition for completing iterations is true, and in the first condition for completing iterations, variables associated with the scaling factor are used, while the first condition for completing iterations is true when the total actual number of bits calculated for all color planes falls within the tolerance range determined at least based on the target size of the output file for JPEG compressed image, or when the width of the first range of the scaling factor is one hundred ovitsya less accuracy parameter scale factor, or when the number of iterations reaches a predetermined maximum number, and if said condition is false, then the scaling factor and said variables are updated, including the boundary values of the first scaling factor range. 2. The method of claim 1, further comprising, before the QT generation step, the steps of: collecting DCT histograms, the DCT histogram collecting step comprising, for each of the color planes, counting the number of occurrences of each DCT coefficient value for each DCT subband; and initialize the target size of the output file for JPEG compressed image, the scaling factor, the boundary values of the range of the first number of bits, the upper and lower tolerances for the first number of bits, the boundary values of the first range of the scaling factor corresponding to the boundary values of the range of the first number of bits, and the accuracy parameter of the scaling factor. 3. The method according to claim 2, in which the step of evaluating the scaling factor comprises the steps of: initialize the boundary values of the range of the second number of bits, the lower and upper tolerances for the second number of bits and the boundary values of the range of the second scaling factor corresponding to the boundary values of the range of the second number of bits; iteratively scaling predetermined base QTs based on the scaling factor to obtain scaled QTs, calculating the total second number of bits as the sum of the second number of bits individually calculated for each of the color planes, the second number of bits being obtained by applying scaled QT to JPEG using the parameters of the prediction models, check whether the second condition for completing iterations is true, and the second condition for completing iterations is true when the full second number of bits falls within the tolerance range for the second number of bits or the width of the second range of the scaling factor becomes less than the accuracy parameter of the scaling factor, and the tolerance range for the second the number of bits is determined based on the target size of the output file and the lower and upper tolerances for the second number of bits, if the second condition for terminating the iterations is false, then the boundary values of the range of the second number of bits, the boundary values of the second range of the scaling factor and the scaling factor are updated based on the full second number of bits using the false position method to find the root. 4. The method of claim 3, wherein said predetermined base QTs are preferably standard JPEG QTs, wherein a second number of bits for a particular color plane is calculated using a polynomial prediction model based on prediction model parameters and prediction parameters for that color plane, wherein the prediction parameters are calculated using the quantized DCT coefficients obtained by quantizing the DCT coefficients using scaled QTs. 5. The method according to claim 4, further comprising, after the step of evaluating the scaling factor, the steps in which: perform quantization and bit counting by using scaled QT in standard JPEG quantization and entropy coding processes, modified so that only the bits output by entropy coding are counted without outputting compressed data, and bit counting is performed separately for each of the color planes and gives as a result, a set of actual number of bits of the color planes, and counting the bits further comprises calculating the total actual number of bits after COROLLARY summing the actual number of bits of color planes for all color planes, however, when verifying the validity of the first condition for completing iterations, the tolerance range is the tolerance range for the first number of bits, which is determined based on the target size of the output file and the lower and upper tolerances for the first number of bits, moreover, at the stage of updating the scaling factor and the corresponding variables, using the false position method to find the root, updating the boundary values of the range of the first number of bits and the boundary values of the first range of the scaling factor based on the total actual number of bits and the current value of the scaling factor and, then, updating the coefficient scaling based on the target size of the output file, updated boundary values of the first coefficient range ienta scaling and updated first boundary amount range of values of bits; wherein the method further comprises, after the step of updating the scaling factor and related variables, the steps in which: updating the prediction model parameter set for each color plane using the set of actual number of bits of the color planes, scaled QTs, and DCT histograms; and if each set of parameters of the prediction model is updated, then the aforementioned predetermined base QTs are scaled based on the scaling factor and return to the step of quantizing and counting bits; otherwise, they return to the step of estimating the scaling factor. 6. The method of claim 5, wherein the step of updating the parameters of the prediction model comprises the steps of: calculating a prediction parameter for each of the color planes using scaled QT and DCT histograms; updating the sets of accumulating variables used for adaptation using the prediction parameter and the normalized actual number of bits of the color plane for the color plane, wherein the sets of accumulating variables are pre-initialized to zero; if the system of linear equations constructed from sets of accumulating variables is joint, then the parameters of the prediction model are calculated by solving the said system of linear equations, otherwise, the parameters of the prediction model are not updated. 7. The method according to claim 6, further comprising stages in which: selecting one of the RC operating modes, the RC operating modes comprising a fast mode providing the best performance; balanced mode that provides a compromise between performance and image quality; and a quality mode that provides the best image quality; calculating the total distortion associated with the scaled QT for each of the color planes; calculate the target number of bits of the color plane for each of the color planes using the set of actual number of bits of the color planes. 8. The method according to claim 7, wherein, if the balanced mode is selected, the method further comprises performing speed-distortion optimization of the balanced mode to generate QT optimized for speed-distortion (RDO QT), wherein the speed-distortion optimization is balanced The mode contains, for each of the color planes, the stages in which: initialize the RDO QT records; calculating the records of the velocity-distortion (RD) tables using DCT histograms and prediction model parameters, said records being dependent on a variable quantizer; determining a set of locally optimal quantizers using RDO QT entries, RD table entries, and a target number of bits of a color plane calculated for said color plane; searching for a locally optimal DCT subband using a set of locally optimal quantizers to determine the index i * of the locally optimal DCT subband; if a locally optimal DCT subband exists, then the quantizer is updated by setting the i * th RDO QT record equal to the i * th locally optimal quantizer and return to the step of determining the set of locally optimal quantizers. 9. The method according to claim 7, in which, if the quality mode is selected, the method further comprises performing speed-distortion optimization of the quality mode for generating an RDO QT, wherein optimizing the speed-distortion of the quality mode comprises, for each of the color planes, the steps , where: calculating RD table entries using DCT histograms and prediction model parameters, said entries being dependent on a variable quantizer; iteratively computes records of optimal quantizer-velocity-distortion (RDQ) tables using RD tables; initialize iterations of obtaining a quantizer based on the records of the RDQ tables and the target number of bits of the color plane calculated for said color plane; iterate the quantizer using the RDQ table entries to create the RDO QT. 10. The method according to p. 8 or 9, further comprising stages in which: calculate the total distortion using RDO QT; perform bit counting using RDO QT in standard JPEG quantization and entropy coding processes, modified so that only bits output during entropy coding are counted without outputting compressed data, and bit counting is performed separately for each of the color planes, and this counting creates the actual number of bits, while counting the bits further comprises calculating the final actual number of bits by summing the actual number of bits for all color planes if the total distortion associated with the RDO QT is less than the total distortion associated with the scaled QT and the final total actual number of bits falls within the tolerance range from the final number of bits, where the tolerance range from the final number of bits is determined based on the target output size file and predefined tolerances for the final number of bits, then output RDO QT as generated QT, otherwise output scaled QT as generated QT T. 11. A computer-readable storage medium on which computer-executable instructions are stored which, when executed by a processor, instruct the processor to execute a speed control (RC) method in JPEG digital image compression, comprising the steps of: receiving discrete cosine transform (DCT) coefficients obtained by applying direct DCT to a digital image; and quantization tables (QT) are generated using the received DCT coefficients by means of iteratively evaluate the scaling factor to create scaled QTs, check whether the first condition for completing iterations is true, and in the first condition for completing iterations, variables associated with the scaling factor are used, while the first condition for completing iterations is true when the total actual number of bits calculated for all color planes falls within the tolerance range determined at least based on the target size of the output file for JPEG compressed image, or when the width of the first range of the scaling factor is one hundred ovitsya less accuracy parameter scale factor, or when the number of iterations reaches a predetermined maximum number, and if said condition is false, then the scaling factor and said variables are updated, including the boundary values of the first scaling factor range. 12. A device for performing speed control (RC) in JPEG compression of a digital image, comprising: one or more processors; and a memory in which computer-executable instructions are stored, which, when executed by one or more processors, prescribe one or more processors: receive discrete cosine transform (DCT) coefficients obtained by applying direct DCT to a digital image; and generate quantization tables (QT) using received DCT coefficients through iterative execution scaling factor estimates to create scaled QTs, checking whether the first condition for completing iterations is true, and in the first condition for completing iterations, variables associated with the scaling factor are used, while the first condition for completing iterations is true when the total actual number of bits calculated for all color planes falls within the range tolerance determined at least based on the target size of the output file for JPEG compressed image, or when the width of the first range of the scaling factor becomes less than the accuracy parameter of the scaling factor, or when the number of iterations reaches a predetermined maximum number, and updating the scaling factor and said variables, including boundary values of the first range of the scaling factor, if said condition is false. 13. A JPEG method for compressing a digital image, comprising the steps of: perform color space conversion and direct discrete cosine transform (DCT); perform a speed control method according to claim 1, perform quantization; perform entropy coding; and output compressed JPEG file.

Images