CN1622135A - Digital image color correction method - Google Patents

Abstract

The present invention relates to method of utilizing nerve network in color correction of digital image. By means of secondary universal rotary combination design, regression equation of hidden layer neure node number and nerve network mean-square deviation is established accurately. The hidden layer neure node number in the nerve network is optimized in genetic algorithm, and this lays the foundation for the determination of hidden layer neure node number. In the required color gamut space, the nerve network is learning trained to obtain the first nerve network with the learning trained color gamut space information, and digital image color information to be corrected is input to the first nerve network to obtain corrected digital image color information. The present invention can correct color of digital image fast precisely and makes it possible to establish nerve network system with operable digital image color correction in great color gamut space.

Description

The bearing calibration of digital image color

Technical field: the invention belongs to intelligence computation and Image Engineering technical field, relate to the method that digital image color is proofreaied and correct.

Background technology: along with various emerging digital color image Input/Output Devices as: the widespread use of color scanner, digital camera, display and color printer, digital color image between various device transmission and the problem of duplicating cross-color in the reproduction process also cause people's attention day by day.This is not only can show the real color of original copy because wish the digital color image that reproduces, and the more important thing is that digital color image is carrying too many information, these informational needs transmit and show with the exact figure color signal, so that the use that the recipient can be correct.The color calibration method of Cai Yonging all was that method is a lot of at local colour gamut or a certain equipment in the past, but its limitation is respectively arranged.Neural network method is a kind of new technology that is used for color correction that grows up over past ten years.

Summary of the invention: determine do not have rationale to seek for the neuron node number that solves each hidden layer of many hidden layers artificial neural network, the blindness that causes the neural network design, color correction will be reached comparatively ideal effect needs a large amount of trials to test to reduce, expending a large amount of time and design repeatedly works and seeks and determine problem such as neuronic node number, and solution expends the problem of a large amount of time trial color reduction, for this reason, the purpose of this invention is to provide a kind of expending time in lacks, accurately determine the neuron node number, the method that obtains the optimum optimization neural network and digital image color is proofreaied and correct.

The present invention is: (1) utilizes the method for the secondary universal rotary combination design in the assay optimization, and the color gamut space of the digital picture of color correction is set up each hidden neuron node of neural network and counted x as required _jRegression equation with neural network square error y: $y=b_0+\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{b}_j{x}_j+\mathrm{\Σ}{b}_{\mathrm{hj}}{x}_h{x}_j+\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{b}_{\mathrm{jj}}{x}_j^2$ (wherein P is the hidden layer number of neural network); Utilize genetic algorithm that regression equation y is optimized again, the minimal value that obtains the neural network square error after the optimization is y _MinThe time to obtain each hidden neuron node number of neural network be x _j(h, j=1 2...P), have obtained each hidden neuron node and have counted x _jJust determined the structure of neural network; The node of color gamut space scope decision hidden neuron is counted x _jScope, the node of hidden neuron is counted x when the color gamut space scope is big _jMany, when the node of color gamut space scope hour hidden neuron is counted x _jFew.

(2) utilize the neural network of above-mentioned optimization,, obtain containing the weight coefficient of required color gamut space colouring information and the first nerves network of threshold value needing the required color gamut space of correction of color that neural network is carried out the weight coefficient training; The colouring information that is corrected is input to the first nerves network, and the colouring information after just can obtaining proofreading and correct has then been finished the correction to digital image color.

The method that a kind of digital image color of the present invention is proofreaied and correct is Function approximation capabilities and the extensive characteristic according to neural network, adopt secondary universal rotary combination design, can set up the relation between variable accurately, compactly, can satisfy actual needs, can reduce test number (TN) again assay optimization.The present invention utilizes genetic algorithm again, employing is based on the heuristic probabilistic global search of the self-adaptation algorithm of the survival of the fittest, the survival of the fittest and colony's theory of evolution of Darwinian evolution, genetic manipulations such as utilization is duplicated, intersected, variation come simulating nature to evolve, and finish the optimizing of each hidden neuron node number of neural network.Be that the present invention utilizes secondary universal rotary combination design and genetic algorithm that neural network is carried out two suboptimization, make the design of neural network form new theoretical system, blindness when having solved the design of background technology neural network, avoided a large amount of trial experiments, not only saved a large amount of time and design repeatedly work, the more important thing is that the design that makes neural network has more science and practicality.Also can set up the neural network that practical exercisable digital image color is proofreaied and correct under the big color gamut space scope.Because the tristimulus value of each pixel of image is all proofreaied and correct by the neural network of optimization of the present invention, solved the problem that expends a large amount of time trial color reduction, the present invention can make the color of digital picture accurately proofread and correct fast.

Description of drawings:

Fig. 1 is a color correction program flow diagram of the present invention

Embodiment: according to the actual alignment requirements of digital image color, use the neural network that the above-mentioned steps method is designed corresponding construction, realize the actual correction of digital image color purpose.Promptly at first carry out secondary universal rotary combination design, obtain each hidden neuron node of neural network and count x _jWith the regression equation of neural network square error y, again this equation is used genetic algorithm optimization, draw its each concrete hidden neuron node and count x _j, in the color gamut space of needs this network is carried out the weight coefficient training, after training finished, this network just can be used for the color correction of above-mentioned training color gamut space.

Embodiments of the invention are as follows:

1. the application of secondary universal rotary combination design

The application of secondary universal rotary combination design is that each hidden neuron of optimal design neural network is counted x _jAnd the regression equation between the neural metwork training error y, factor adopts the neural network hidden layer number of P 〉=2, at first factor is encoded, encode from the natural cause space to the coding factor Spaces, select the corresponding composite design of coding factor Spaces for use, assortment testing program and computation scheme table, calculating and statistical test through regression coefficient, just obtain the regression equation of the factor of encoding, natural cause is brought into the regression equation of coding factor, just natural cause regression equation and this equation substitution genetic algorithm is optimized, try to achieve the best neuron number x of each hidden layer _jDuring as P=2, each hidden neuron node of network is counted x _jWith the regression equation of neural network square error y be:

$y=b_0+\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{b}_j{x}_j+\mathrm{\Σ}{b}_{\mathrm{hj}}{x}_h{x}_j+\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{b}_{\mathrm{jj}}{x}_j^2=b_0+b_1{x}_1+b_2{x}_2+b_{12}{x}_1{x}_2+b_{11}{x}_1^2+b_{22}{x}_2^2$

$(h,j=\mathrm{1,2}...P).$

2. utilize the genetic algorithm for solving process

Genetic algorithm is the heuristic probabilistic global search of the self-adaptation algorithm of a kind of survival of the fittest based on Darwinian evolution, the survival of the fittest and colony's theory of evolution.Genetic manipulations such as the genetic algorithm utilization is duplicated, intersected, variation come simulating nature to evolve, and finish the problem optimizing, and its basic step is as follows:

(1) coding.Independent variable to each problem to be optimized need be encoded, the binary code of general employing finite length is represented the various values of independent variable, if the binary code of each independent variable is strung, obtain a binary code string, on behalf of one group of that value determined of independent variable, it separate.As each being separated the body of regarding as in the biotic population one by one, above-mentioned code then is equivalent to represent the chromosome of this individual inheritance characteristic.

(2) produce initial population.Produce n bar chromosome at random and form initial population.This population is represented the set of some feasible solutions of optimization problem, and generally speaking the quality of initial population is relatively poor, and genetic algorithm is from this initial population, and the simulated evolution process is rogued according to qualifications, finds out outstanding colony and individuality at last, satisfies the requirement of optimizing.

(3) duplicate.Press coding rule, the regression equation with the pairing independent variable value of each the individual chromosome substitution objective function expression formula y in the colony calculates adaptive value.From colony, choose M to individuality by certain probability, be replicated, be used to raise up seed as parents, duplicating principle is the bigger individuality of adaptive value, gives the bigger probability of choosing, therefore, the adaptive value high individuality of healing has more opportunity to raise up seed, and makes its good characteristic be able to heredity and reservation.

(4) intersect.The parents of selected at random are matched arbitrarily, respectively it is intersected then.The simplest intersection way is to choose one or more truncation points randomly, and parents' chromosome is separated at truncation points, then, exchanges its afterbody with certain probability.

(5) variation.At first by given several individualities of probability picked at random.Generally speaking, given variation probability is all very little, generally gets 0～0.05.To each individuality of having chosen, the negate computing is carried out in a certain position of picked at random, promptly by 1 → 0 or by 0 → 1.

(6) produced population of new generation by above-mentioned (3)～(5) after, each individuality of new colony is estimated again, promptly each chromosome is decoded, and obtains each individual adaptive value.

(7) repeat above-mentioned steps, reach a certain set value up to the adaptive value of optimum individual, or the average adaptive value of the adaptive value of optimum individual and population no longer improves, then iterative process finishes, and draws the best neuron number x of each hidden layer _j

3. the application of neural network

The problem that neural network can not be described with rule or formula handling a large amount of raw data, or when the rules such as mechanism of problem are had little understanding, all show great dirigibility and adaptivity.The error backpropagation algorithm of neural network promptly is made of the forward calculation (forward-propagating) of data stream and two processes of backpropagation of error signal.During forward-propagating, the direction of propagation is input layer → hidden layer → output layer, and every layer of neuronic state only influences one deck neuron down.As if the output that can not get expecting at output layer, then the backpropagation flow process of steering error signal.By hocketing of these two processes, carry out error function gradient decline strategy in the weight vector space, dynamically one group of weight vector of iterative search makes the network error function reach minimum value, thereby makes neural network finish information extraction and memory process.In the required color space, provide one group of learning data, neural network to be trained, after neural metwork training finished, the rule when this neural network just can be according to learning training was used for the color correction in required color space.

4. any width of cloth digital picture is written into computing machine, presses Fig. 1 flow processing digital picture, just can obtain the digital picture of a width of cloth color through overcorrect, and its trimming process is as follows:

The digital picture that a need proofread and correct is read in internal memory; The definition of data pointer also points to first pixel of Digital Image Data;

Whether b judgment data pointer points to the last pixel of Digital Image Data;

C does not point to the last pixel of Digital Image Data, tristimulus value to first pixel of digital picture is proofreaied and correct by the first nerves network that trains, tristimulus value after the correction generates a new pixel, is written in the corresponding pixel of the new digital image file position;

The d data pointer points to the next pixel of digital picture;

E repeats (2)-(5) step, successively the tristimulus value of each pixel in the digital picture is proofreaied and correct by the first nerves network

The f data pointer points to digital picture end pixel, and the whole pixel corrections of digital picture are finished, and generate new digital picture.

Claims (2)

1, the bearing calibration of digital image color is characterized in that:

(1) utilize the method for secondary universal rotary combination design in the assay optimization, the color gamut space of the digital picture of color correction is set up the regression equation that each hidden neuron node of neural network is counted xj and neural network square error y as required: $y=b_0+\underset{i=1}{\overset{p}{\mathrm{\Σ}}}{b}_j{x}_j+\mathrm{\Σ}{b}_{\mathrm{hj}}{x}_h{x}_j+\underset{j=1}{\overset{p}{\mathrm{\Σ}}}{b}_{\mathrm{jj}}{x}_j^2;$ Utilize genetic algorithm that regression equation y is optimized again, the minimal value that obtains the neural network square error after the optimization is y _MinThe time to obtain each hidden neuron node number of neural network be x _j, obtained each hidden neuron node and counted x _jJust determined the structure of neural network;

(2) utilize the neural network of optimizing,, obtain containing the weight coefficient of required color gamut space colouring information and the first nerves network of threshold value needing the required color gamut space of correction of color that above-mentioned neural network is carried out the weight coefficient training; The colouring information that needs are proofreaied and correct is input to the first nerves network, and the whole colouring informations after obtaining proofreading and correct have then been finished the correction to digital image color.

2, the bearing calibration of digital image color according to claim 1 is characterized in that:

The digital picture that a need proofread and correct is read in internal memory, and the definition of data pointer also points to first pixel of Digital Image Data;

Whether b judgment data pointer points to the last pixel of Digital Image Data;

C does not point to the last pixel of Digital Image Data, tristimulus value to first pixel of digital picture is proofreaied and correct by the first nerves network that trains, tristimulus value after the correction generates a new pixel, is written in the corresponding pixel of the new digital image file position;

The d data pointer points to the next pixel of digital picture;

E repeats (2)-(5) step, successively the tristimulus value of each pixel in the digital picture is proofreaied and correct by the first nerves network;

The f data pointer points to digital picture end pixel, and the whole pixel corrections of digital picture are finished, and generate new digital picture.