CN102111626A - Conversion method and device from red-green-blue (RGB) color space to cyan-magenta-yellow-black (CMYK) color space - Google Patents

Abstract

The invention provides a conversion method and device from a red-green-blue (RGB) color space to a cyan-magenta-yellow-black (CMYK) color space. The method comprises the steps of model configuration, data acquisition and model learning, wherein in the step of model configuration, a neural network mapping model is configured, R, G and B variables in the RGB space are input by the mapping model, and C, M, Y and K variables in the CMYK space output by the mapping model are taken as the output variables of the conversion model, and the parameters of the mapping model comprise a network weight and a network threshold; in the step of data acquisition, N RGB color values and corresponding CMYK color values are used as the input and desired output of a learning sample, so as to obtain the learning sample; and in the step of model learning, aiming at the learning sample, the mapping model is trained by using a back propagation (BP) algorithm so as to obtain the model parameters, thereby determining the mapping model. According to the invention, the conversion error from the RGB color space to the CMYK color space is reduced.

Description

A kind of RGB is to the conversion method and the device of cmyk color space

Technical field

The present invention relates to technical field of image processing, particularly relate to conversion method and the device of a kind of RGB to cmyk color space.

Background technology

The color transformed colour space transformation that is meant of broad sense, what be about to also that a certain color space represents color transformedly represents in another kind of color space, as common RGB (red Red, green Green, blue Blue), CMYK (blue or green Cyan, fuchsin Magenta, yellow Yellow, black Black), the mutual conversion in Lab color spaces such as (illuminance Luminosity, a, b represent three-dimensional two axles of color respectively).

The most frequently used purposes of rgb color space is exactly a display system, the display of color cathode ray tube, chromatic grating figure all uses R, G, B numerical value to drive R, G, B electron gun emitting electrons, and R, the G on the fluorescence excitation screen, the fluorescent material of three kinds of colors of B send the light of different brightness respectively, and mix by addition and to produce shades of colour; Scanner also is by absorbing R, G, the B composition of original copy in the light that reflection or transmission are sent, and represents the color of original copy with it.Rgb color space is called device-dependent color space, because the same width of cloth image of different scanner scanning, can obtain the view data of different color; The display of different model shows same width of cloth image, also has different color display result.

Cmyk color space is applied to printing industry, and press shows colourful color and contrast by the double exposure of the different dot area percentages of blue or green (C), product (M), yellow (Y) three-colour printing ink, and this is trichromatic CMY color space.In the actual print, generally adopt blue or green (C), product (M), yellow (Y), black (K) four-color process, increase black print in middle tone to the shadow of printing.When the RGB three primary colors are mixed, can produce white, but when mixing blue-green, aubergine and yellow three primary colors, can produce black.Since the ink of actual usefulness can't produce pure color, black is to be included in color separately, and this model is referred to as CMYK.Cmyk color space is relevant with equipment or printing process, and then for the characteristic of process, printing ink, the conditions such as characteristic of paper, different conditions has different print result.So the CMYK color space is called the table color space relevant with equipment.

And CMYK has ambiguity, that is to say with a kind of color with identical Absolute Colorimetric, under identical printing process prerequisite, can represent and prints with minute kind CMYK digit groups is incompatible.This specific character has been brought a lot of troubles to color management, has equally also brought a lot of flexibilities to control.In printing process, must be through the process of a color separation, so-called color separation is exactly that the RGB color conversion that will use in the computer becomes to print the CMYK color of using.

Exist two complicated problems in transfer process, these two color spaces of the first are not exclusively the same on the scope of apparent color, and the colour gamut of RGB is bigger, and CMYK is then less, therefore will carry out gamut compression; It two is that these two colors all are that itself does not have absoluteness color with concrete device-dependent.Therefore, existing conversion method is often changed with device-independent color space by one.For example, as a kind of inner color mode, also, the execution sequence of existing conversion method is RGB → Lab at first, is only Lab → CMYK then with the LAB pattern.Above-mentioned transformational relation nonlinearity height, transformed error is often not ideal enough.

In a word, need the urgent technical problem that solves of those skilled in the art to be exactly: how can solve RGB to the big problem of the transformed error of cmyk color space.

Summary of the invention

Technical problem to be solved by this invention provides conversion method and the device of a kind of RGB to cmyk color space, in order to reduce the transformed error of RGB to cmyk color space.

In order to address the above problem, the invention discloses the conversion method of a kind of RGB to cmyk color space, comprising:

Model construction step: constructing neural network mapping model, the input of this mapping model is R, G, the B variable in the rgb space, output is C, M, Y, the K variable in the CMYK space, and as the output variable of this transformation model, the parameter of this mapping model comprises network weight and network threshold;

Data acquisition step: with N RGB color value and corresponding CMYK color value is the input and the desired output of learning sample, obtains learning sample;

Model learning step: at described learning sample, adopt the BP algorithm to train this mapping model, obtain model parameter, thereby determine this mapping model.

Preferably, the structure of described mapping model comprises input layer, hidden layer and output layer, and wherein, the input layer number is 3, and output layer node number is 4, and the number of hidden nodes is

Wherein, 1＜a＜30.

Preferably, described model learning step comprises:

The forward-propagating process:

Each neuron of input layer receives the learning sample information of input, and passes to each neuron of hidden layer;

Hidden layer is carried out conversion to described learning sample information, and passes to each neuron of output layer;

Output layer carries out conversion to the output information of described hidden layer, obtains actual output;

Deterministic process:

Calculate the error between actual output and the desired output, and judge whether this error satisfies predictive error, if then algorithm finishes; Otherwise, carry out back-propagation process;

Back-propagation process:

Described error signal by output layer, to the anti-pass successively of hidden layer, input layer, wherein, is pressed each layer of mode correction weights that error gradient descends, and when propagating end, returned the forward-propagating process in the anti-pass process.

Preferably, described model learning step comprises:

Initialization procedure: preset study number of times T, set current study number of times t=0,

The forward-propagating process:

For current study number of times t, each neuron of input layer receives the learning sample information of input, and passes to each neuron of hidden layer;

Hidden layer is carried out conversion to described learning sample information, and passes to each neuron of output layer;

Output layer carries out conversion to described hidden layer output information, obtains actual output;

Back-propagation process:

Described error signal by output layer, to the anti-pass successively of hidden layer, input layer, wherein, is pressed each layer of mode correction weights that error gradient descends in the anti-pass process;

Deterministic process:

Judge whether t＞T-2 sets up, if then algorithm finishes; Otherwise, upgrade study number of times t=t+1, and return the forward-propagating process.

Preferably, the transfer function of described hidden layer and output layer be Sigmoid type: f (x)=1/[1+e^ (bx)], b＞0.

Preferably, when b=1, described data acquisition step comprises:

Obtain N RGB color value and corresponding CMYK color value, and with its input and desired output initial data as learning sample;

Normalized is carried out in described input and desired output initial data, make its value between [0,1].

Preferably, described normalization step is using formula

Described initial data is carried out normalized, make its value between [0,1], wherein, x is an initial data, and X ' is a normalization data.

The invention also discloses the conversion equipment of a kind of RGB, comprising to cmyk color space:

The model construction module, be used for the constructing neural network mapping model, the input of this mapping model is R, G, the B variable in the rgb space, and output is C, M, Y, the K variable in the CMYK space, as the output variable of this mapping model, the parameter of this mapping model comprises network weight and network threshold;

Data acquisition module, being used for N RGB color value and corresponding CMYK color value is the input and the desired output of learning sample, obtains learning sample;

The model learning module is used at described learning sample, adopts the BP algorithm to train this mapping model, obtains model parameter, thereby determines this mapping model.

Preferably, described model learning module comprises:

The forward-propagating submodule comprises:

The input layer unit is used to receive the learning sample information of input, and passes to each neuron of hidden layer;

The hidden layer unit is used for described learning sample information is carried out conversion, and passes to each neuron of output layer;

Output layer unit is used for the output information of described hidden layer is carried out conversion, obtains actual output;

Judge submodule, be used to calculate the error between actual output and the desired output, and judge whether this error satisfies predictive error, if then algorithm finishes; Otherwise, carry out back-propagation process; And

The backpropagation submodule is used for described error signal is passed through output layer, to the anti-pass successively of hidden layer, input layer, wherein, each layer of the mode correction weights that descend by error gradient in the anti-pass process, and when propagating end, triggering forward-propagating submodule.

Preferably, the structure of described mapping model comprises input layer, hidden layer and output layer, the transfer function of described hidden layer and output layer be Sigmoid type: f (x)=1/[1+e^ (x)];

Described acquisition module comprises:

Color value obtains submodule, is used to obtain N CMYK color value and corresponding RGB color value, and with its input and desired output initial data as learning sample;

The normalization submodule is used for normalized is carried out in described input and desired output initial data, makes its value between [0,1].

Compared with prior art, the present invention has the following advantages:

The present invention utilizes the BP neural net to construct the mapping model of RGB to cmyk color space, because a large amount of input-output mapping relations can be learnt and store to the BP neural net, and need not to disclose in advance the math equation of describing this mapping relations; In addition, the learning process of this mapping model, it is the process that model parameter is constantly revised, the error that this process can be performed until mapping model output reduces to predetermined accuracy, thereby, the present invention can finish the approximate accurately mapping of 3 dimension rgb color space to 4 dimension cmyk color spaces under the prerequisite of little error.

Description of drawings

Fig. 1 is the flow chart of a kind of RGB of the present invention to the conversion method embodiment of cmyk color space;

Fig. 2 is the topology diagram of a kind of BP neural net of the present invention;

Fig. 3 is the neuronic schematic diagram of a kind of basic BP of the present invention;

Fig. 4 is a kind of sigmoid examples of functions of the present invention;

Fig. 5 is the flow chart that a kind of error of the present invention is adjusted scheme.

Embodiment

For above-mentioned purpose of the present invention, feature and advantage can be become apparent more, the present invention is further detailed explanation below in conjunction with the drawings and specific embodiments.

Because the different table color method has different colour gamuts and precision in the space information system, usually can cause this transformation relation can not be with simple explicit the expression, thereby cause color transformed relation very complicated.

The rgb color pattern uses the RGB model to distribute one 0～255 intensity level in the scope as R, G, the B component of each pixel in the image.For example: pure red R value is 255, and the G value is 0, and the B value is 0; The R of grey, G, three values of B equate (except 0 and 255); R, G, the B of white are 255; The R of black, G, B are 0.The RGB image only uses three kinds of colors, just can make them according to different mixed, reappears 16777216 kinds of colors on screen.

And in the image of cmyk color pattern, each pixel all is by C, M, Y and K look synthetic according to different ratios.Every kind of printing-ink of each pixel can be assigned with a percent value, the printing ink colors percent value that the color assignment of the brightest (Gao Guang) is lower, the higher percent value of color assignment of dark (shadow).For example, Ming Liang redness may comprise 2% cyan, 93% fuchsin, 90% yellow and 0% black.In the CMYK image, when the value of all 4 kinds of components all is 0%, will produce pure white.

One of core idea of the present invention is, utilize the BP neural net to construct the mapping model of RGB to cmyk color space, because a large amount of input-output mapping relations can be learnt and store to the BP neural net, and need not to disclose in advance the math equation of describing this mapping relations; Its learning rules are to use the gradient descent method, constantly adjust the weights and the threshold value of network by backpropagation, make the error minimum of network, thereby can solve RGB of the present invention to the big problem of the transformed error of cmyk color space.

With reference to Fig. 1, show the flow chart of a kind of RGB of the present invention to the conversion method embodiment of cmyk color space, specifically can comprise:

Model construction step 101, constructing neural network mapping model, the input of this mapping model can be the R in the rgb space, G, B variable, output can be the C in the CMYK space, M, Y, K variable, as the output variable of this transformation model, the parameter of this mapping model can comprise network weight and network threshold;

BP (Back Propagation) neural net is a kind of Learning Algorithm, and full name is based on the artificial neural net of error backpropagation algorithm.With reference to topological structure shown in Figure 2, the BP neural network model generally comprises input layer (input), hidden layer (hide layer) and output layer (output layer).Its feature is: each layer neuron only with between the adjacent layer neuron is connected mutually entirely, do not connect with having between the neuron in the layer, do not have feedback between each layer neuron and connects, and constitutes the feed-forward type nerve network system with hierarchical structure.

With reference to figure 3, show j basic BP neuron, it has imitated three that biological neuron had substantially the most also is most important function: weighting, summation and transfer.Wherein, x1, x2...xi...xn represent the input from neuron 1,2...i...n respectively; Wj1, wj2...wji...wjn then represent neuron 1,2...i...n and j neuronic bonding strength, i.e. weights respectively; Bj is a threshold value; F () is a transfer function; Yj is j neuronic output.

J neuronic clean input value S _jFor:

$S_j=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}\·x_i+b_j=W_{j}X+b_j$

Wherein, X=[x ₁x ₂X _iX _n] ^TW _j=[w _J1w _J2W _JiW _Jn]

If look x ₀=1, w _J0=b _jEven, X and W _jComprise x ₀And w _J0, then

X＝[x ₀?x ₁?x ₂…x _i…x _n] ^T?W _j＝[w _j0?w _j1?w _j2…w _ji…w _jn]

So clean input S of node j _jCan be expressed as:

$S_j=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}{x}_i=W_{j}X$

Clean input S _jBehind transfer function (Transfer Function) f (), just obtain j neuronic output y _j:

$y_j=f\left(s_j\right)=f(\underset{i=0}{\overset{n}{\mathrm{\Σ}}}{w}_{\mathrm{ji}}\·x_i)=F\left(W_{j}X\right)$

F () is dull increasing function in the formula, and must be bounded function, because the signal that cell transmits can not infinitely increase.

To sum up, determined that the network number of plies, every node layer number, transfer function, initial weight coefficient etc. have also just determined the BP network.Certain guideline is arranged when determining these options, but more be by experience.

Specific to the embodiment of the invention, then the structure of described transformation model can comprise input layer, hidden layer and output layer, and the input layer number is 3, and output layer node number is 4.

For Multi-layered Feedforward Networks, determining of the number of hidden nodes is key of success.If quantity is very little, then network can obtain in order to the information of dealing with problems very little; If quantity is too many, not only increase the training time, and (Overfitting) problem of what is called " transition coincide " too much also may appear in hidden node, i.e. test error increase causes generalization ability to descend, and therefore, the choose reasonable the number of hidden nodes is extremely important.

About the selection more complicated of hidden layer number and node number thereof, rule is: correctly reflecting on the basis of input/output relation, should select less the number of hidden nodes for use, so that network configuration is simple as far as possible.Therefore, the embodiment of the invention is preferentially selected single hidden layer structure for use, and selects the number of hidden nodes according to following empirical equation:

Wherein, 1＜a＜30.

The transfer function of BP network has multiple.For example, Matlab (the matrix experiment chamber provides following transfer function in MatrixLaboratory):

1. the pure linear shape function of PURELIN

Form: PU RELIN (Z, B), can the batch processing vector Z, and threshold value B is provided;

2. logarithm S shape LOGSIG function

Form: LOGSIG (Z, B);

3. tanh S shape TANSIG function

Form: TANSIG (Z, B).

In a preferred embodiment of the present invention, adopt sigmoid type function f (x)=1/[1+e^ (bx)] (b＞0) as the transfer function of hidden layer and output layer.With reference to figure 4, show the sigmoid examples of functions of b=1, it is a good threshold function table, has continuously, smooth, strictly monotone is about (0,0.5) centrosymmetric characteristics.

Data acquisition step 102, be the input and the desired output of learning sample, obtain learning sample with N RGB color value and corresponding CMYK color value;

For example,, be in the transfer function of hidden layer and output layer for the forecast model of single hidden layer: f (x)=1/[1+e^ (x)] time, its output valve is between 0 and 1, and at this moment, described obtaining step can comprise following substep:

Substep A1, obtain N RGB color value and corresponding CMYK color value, and with its input and desired output initial data as learning sample;

Substep A2, normalized is carried out in described input and desired output initial data, make its value between [0,1].

For example, described normalization step can be, using formula

Described initial data is carried out normalized, make its value between [0,1], wherein, x is an initial data, and X ' is a normalization data.

Model learning step 103, at described learning sample, adopt the BP algorithm to train this mapping model, obtain model parameter, thereby determine this mapping model.

The learning process of BP algorithm is made up of the forward-propagating of signal and two processes of backpropagation of error.

1) forward-propagating: input sample-input layer-each hidden layer (processing)-output layer;

2) error back propagation: output error (certain form)-hidden layer (successively)-input layer;

Its main purpose is by with the output error anti-pass, error is shared to all unit of each layer, thereby obtain the error signal of each layer unit, and then revise the weights (its process is the process that weights are adjusted) of each unit.

Therefore, information forward-propagating that goes round and begins again and error back propagation process are the processes that each layer weights are constantly adjusted, and also are the processes of neural network learning training, the error that this process is performed until network output reduces to the acceptable degree, till the perhaps predefined study number of times.

Correspondingly, the embodiment of the invention can comprise following two kinds of study schemes:

One, error is adjusted scheme

Main thought can be understood as: to the network configuration of determining, according to the size of error between desired output and the real output value, adjust weights and threshold value repeatedly, till this error reaches predetermined value.Particularly, can be when the output that output layer can not get expecting, the back-propagation process of turning error signal specifically can comprise following process:

Forward-propagating process A1,

Each neuron of input layer receives the learning sample information of input, and passes to each neuron of hidden layer;

Hidden layer is carried out conversion to described learning sample information, and passes to each neuron of output layer;

Output layer carries out conversion to the output information of described hidden layer, obtains actual output;

Deterministic process A2,

Calculate the error between actual output and the desired output, and judge whether this error satisfies predictive error, if then algorithm finishes; Otherwise, carry out back-propagation process;

Back-propagation process A3,

Described error signal by output layer, to the anti-pass successively of hidden layer, input layer, wherein, is pressed each layer of mode correction weights that error gradient descends, and when propagating end, returned A1 in the anti-pass process.

For making those skilled in the art understand the present invention better, below this programme is described by concrete example.

The learning sample obtaining step of supposing this example is to obtain N=625 CMYK color value and corresponding RGB color value, as learning sample;

Transformation model is three layers of BP network of a single hidden layer:

The input layer number is 3, establishes Y _i ¹Output for input layer i (i=0,1,2);

The number of hidden nodes n ₁=18, the characteristic of each node be Sigmoid type: f (x)=1/[1+e^ (x)], Y _j ²For middle layer node j (j=0,1,2 ..., 16,17) output;

Output layer node number is 4, the characteristic of each node be Sigmoid type: f (x)=1/[1+e^ (x)], Y _k ³Be the output of output layer node k (k=0,1,2,3), T _kBe the corresponding desired output of output layer node k (k=0,1,2,3);

W _IjBe the connection weights between node i and the node j, W _JkBe the connection weights between node j and the node k, θ _jBe the threshold value of middle layer node j, θ _kThreshold value for output layer node k;

Evaluation function is square type error function

Learning criterion adopts the BP algorithm also promptly, the gradient descent method, and it carries out error function gradient decline strategy in the weight vector space, and dynamically one group of weight vector of iterative search makes the network error function reach minimum value, thereby finishes information extraction and memory process.

With reference to Fig. 5, show the realization flow figure of this programme, specifically can comprise:

Initialization operation 501, preset number of samples p=0, the default local error upper limit and the global error upper limit, network weight and network threshold are carried out initialization, wherein, described network weight can comprise the connection weights between output layer node and the hidden node, and the connection weights between hidden node and the input layer, described network threshold can comprise hidden node threshold value and output layer node threshold value;

For example, give little random number, W to network weight and threshold value _Ij(t) ∈ [1,1], W _Jk(t) ∈ [1,1], θ _j(t) ∈ [1,1], θ _k(t) ∈ [1,1];

Input operation 502, p learning sample of input, as current learning sample, wherein, p ∈ 1,2 ..., N};

The output valve of first calculating operation 503, each node of calculating hidden layer $Y_j^{}$ ${}_2^{}=f(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{W}_{\mathrm{ij}}{Y}_i^1-{\mathrm{\θ}}_j)=f(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{W}_{\mathrm{ij}}{X}_k-{\mathrm{\θ}}_j);$

The output valve of second calculating operation 504, each node of calculating output layer

The 3rd calculating operation 505, at current sample, based on the output valve of desired output and each node of output layer, adopt square type error function, calculate the error of current sample

First decision operation 506, judge current sample error whether less than the local error upper limit, if then carry out second decision operation 507; Otherwise, carry out first and revise operation 510;

Second decision operation 507, judge whether K＞N-1 sets up, if then carry out the 4th calculating operation 508; Otherwise, carry out first and revise operation 510;

The 4th calculating operation 508, at all N sample, based on the output valve of desired output and each node of output layer, adopt square type error function, calculate global error

The 3rd decision operation 509, whether judge global error, if then algorithm finishes less than the global error upper limit; Otherwise, carry out first and revise operation 510;

First revises the connection weights correction between operation 510, calculating output layer node and the hidden node:

And according to calibration corrections δ _kRevise connection weight value matrix W between output layer and hidden layer _JkWith threshold vector θ _k

For example, node K and hidden layer j be connected weights W _JkWith being modified to of the threshold value of node K:

θ _k(t+1)=θ _k(t)+β δ _k

Second revises the connection weights correction between operation 511, calculating hidden node and the input layer: And according to described calibration corrections δ _jRevise the connection weight value matrix W between hidden layer and input layer _JiWith threshold vector θ _j, and make K=K+1, return input operation 502.

For example, hidden layer j and input layer i's is connected weights W _JiThreshold vector θ with node j _jCorrection value be:

θ _j(t+1)=θ _j(t)+α δ _j

Above-mentioned example adopts global error as interpretational criteria, is appreciated that those skilled in the art can also adopt other interpretational criterias as required, as the employing local error, and mean square error (MSE, MeanSquare Error) etc., the present invention is not limited this.

Two, the study number of times is adjusted scheme

The study number of times is an important parameter of neural net.The study number of times too much can cause the mistake of network to fit, and causes the result to produce deviation; The study number of times is crossed and is made network be difficult to convergence at least, does not reach study requirement.The thought of this programme is, finishes learning process by predefined study number of times.

In specific implementation, this scheme can comprise following process:

Initialization procedure B1, preset study number of times T, set current study number of times t=0,

Forward-propagating process B2,

For current study number of times t, each neuron of input layer receives the learning sample information of input, and passes to each neuron of hidden layer;

Hidden layer is carried out conversion to described learning sample information, and passes to each neuron of output layer;

Output layer carries out conversion to described hidden layer output information, obtains actual output;

Back-propagation process B3,

Described error signal by output layer, to the anti-pass successively of hidden layer, input layer, wherein, is pressed each layer of mode correction weights that error gradient descends in the anti-pass process;

Deterministic process B4,

Judge whether t＞T-2 sets up, if then algorithm finishes; Otherwise, upgrade study number of times t=t+1, and return B2.

Be appreciated that, the present invention is not limited to the back propagation learning rule of BP algorithm, can also be applicable to other learning rules, as utilizing the improvement BP algorithm of momentum rule, self adaptation to adjust learning rate, momentum-adaptive learning speed adjustment algorithm, L-M (Levenberg-Marquardt) learning rules etc., those skilled in the art can select as required.

In addition, above-mentioned two kinds of schemes can also be combined, weights and threshold value are adjusted, the present invention is not limited this.

The invention also discloses the conversion equipment embodiment of a kind of RGB, specifically can comprise to cmyk color space:

The model construction module, be used for the constructing neural network mapping model, the input of this mapping model can be the R in the rgb space, G, B variable, output can be the C in the CMYK space, M, Y, K variable, as the output variable of this mapping model, the parameter of this mapping model can comprise network weight and network threshold;

Data acquisition module, being used for N RGB color value and corresponding CMYK color value is the input and the desired output of learning sample, obtains learning sample;

For the embodiment of the invention, the structure of described mapping model can be a three-decker, also, can comprise input layer, hidden layer and output layer, the transfer function of described hidden layer and output layer can for Sigmoid type: f (x)=1/[1+e^ (x)];

At this moment, described acquisition module can comprise:

Color value obtains submodule, is used to obtain N CMYK color value and corresponding RGB color value, and with its input and desired output initial data as learning sample;

The normalization submodule is used for normalized is carried out in described input and desired output initial data, makes its value between [0,1];

The model learning module is used at described learning sample, adopts the BP algorithm to train this mapping model, obtains model parameter, thereby determines this mapping model.

In specific implementation, can adopt error to adjust scheme this mapping model is learnt.Also promptly, to the network configuration of determining, according to the size of error between desired output and the real output value, adjust weights and threshold value repeatedly, till this error reaches predetermined value.Correspondingly, the structure of described model learning module can for:

The forward-propagating submodule specifically can comprise:

The input layer unit is used to receive the learning sample information of input, and passes to each neuron of hidden layer;

The hidden layer unit is used for described learning sample information is carried out conversion, and passes to each neuron of output layer;

Output layer unit is used for the output information of described hidden layer is carried out conversion, obtains actual output;

Judge submodule, be used to calculate the error between actual output and the desired output, and judge whether this error satisfies predictive error, if then algorithm finishes; Otherwise, carry out back-propagation process; And

The backpropagation submodule is used for described error signal is passed through output layer, to the anti-pass successively of hidden layer, input layer, wherein, each layer of the mode correction weights that descend by error gradient in the anti-pass process, and when propagating end, triggering forward-propagating submodule.

Certainly, those skilled in the art can also be as required, adopts the study number of times to adjust scheme, also, finishes learning process by predefined study number of times, and the structure of described model learning module is designed, and the present invention is not limited this.

For device embodiment, because it is similar substantially to method embodiment shown in Figure 1, so description is fairly simple, relevant part gets final product referring to the part explanation of method embodiment.

More than to conversion method and the device of a kind of RGB provided by the present invention to cmyk color space, be described in detail, used specific case herein principle of the present invention and execution mode are set forth, the explanation of above embodiment just is used for helping to understand method of the present invention and core concept thereof; Simultaneously, for one of ordinary skill in the art, according to thought of the present invention, the part that all can change in specific embodiments and applications, in sum, this description should not be construed as limitation of the present invention.

Claims (10)

1. a RGB is characterized in that to the conversion method of cmyk color space, comprising:

Model construction step: constructing neural network mapping model, the input of this mapping model is R, G, the B variable in the rgb space, output is C, M, Y, the K variable in the CMYK space, and as the output variable of this transformation model, the parameter of this mapping model comprises network weight and network threshold;

Data acquisition step: with N RGB color value and corresponding CMYK color value is the input and the desired output of learning sample, obtains learning sample;

Model learning step: at described learning sample, adopt the BP algorithm to train this mapping model, obtain model parameter, thereby determine this mapping model.

2. the method for claim 1 is characterized in that, the structure of described mapping model comprises input layer, hidden layer and output layer, and wherein, the input layer number is 3, and output layer node number is 4, and the number of hidden nodes is Wherein, 1＜a＜30.

3. method as claimed in claim 2 is characterized in that, described model learning step comprises: the forward-propagating process:

Each neuron of input layer receives the learning sample information of input, and passes to each neuron of hidden layer;

Hidden layer is carried out conversion to described learning sample information, and passes to each neuron of output layer;

Output layer carries out conversion to the output information of described hidden layer, obtains actual output;

Deterministic process:

Calculate the error between actual output and the desired output, and judge whether this error satisfies predictive error, if then algorithm finishes; Otherwise, carry out back-propagation process;

Back-propagation process:

Described error signal by output layer, to the anti-pass successively of hidden layer, input layer, wherein, is pressed each layer of mode correction weights that error gradient descends, and when propagating end, returned the forward-propagating process in the anti-pass process.

4. method as claimed in claim 2 is characterized in that, described model learning step comprises:

Initialization procedure: preset study number of times T, set current study number of times t=0,

The forward-propagating process:

For current study number of times t, each neuron of input layer receives the learning sample information of input, and passes to each neuron of hidden layer;

Hidden layer is carried out conversion to described learning sample information, and passes to each neuron of output layer;

Output layer carries out conversion to described hidden layer output information, obtains actual output;

Back-propagation process:

Described error signal by output layer, to the anti-pass successively of hidden layer, input layer, wherein, is pressed each layer of mode correction weights that error gradient descends in the anti-pass process;

Deterministic process:

Judge whether t＞T-2 sets up, if then algorithm finishes; Otherwise, upgrade study number of times t=t+1, and return the forward-propagating process.

5. method as claimed in claim 2 is characterized in that, the transfer function of described hidden layer and output layer be Sigmoid type: f (x)=1/[1+e^ (bx)], b＞0.

6. method as claimed in claim 5 is characterized in that, when b=1, described data acquisition step comprises:

Obtain N RGB color value and corresponding CMYK color value, and with its input and desired output initial data as learning sample;

Normalized is carried out in described input and desired output initial data, make its value between [0,1].

7. method as claimed in claim 6 is characterized in that, described normalization step is using formula

Described initial data is carried out normalized, make its value between [0,1], wherein, x is an initial data, and X ' is a normalization data.

8. a RGB is characterized in that to the conversion equipment of cmyk color space, comprising:

The model construction module, be used for the constructing neural network mapping model, the input of this mapping model is R, G, the B variable in the rgb space, and output is C, M, Y, the K variable in the CMYK space, as the output variable of this mapping model, the parameter of this mapping model comprises network weight and network threshold;

Data acquisition module, being used for N RGB color value and corresponding CMYK color value is the input and the desired output of learning sample, obtains learning sample;

The model learning module is used at described learning sample, adopts the BP algorithm to train this mapping model, obtains model parameter, thereby determines this mapping model.

9. device as claimed in claim 8 is characterized in that, described model learning module comprises:

The forward-propagating submodule comprises:

The input layer unit is used to receive the learning sample information of input, and passes to each neuron of hidden layer;

The hidden layer unit is used for described learning sample information is carried out conversion, and passes to each neuron of output layer;

Output layer unit is used for the output information of described hidden layer is carried out conversion, obtains actual output;

Judge submodule, be used to calculate the error between actual output and the desired output, and judge whether this error satisfies predictive error, if then algorithm finishes; Otherwise, carry out back-propagation process; And

The backpropagation submodule is used for described error signal is passed through output layer, to the anti-pass successively of hidden layer, input layer, wherein, each layer of the mode correction weights that descend by error gradient in the anti-pass process, and when propagating end, triggering forward-propagating submodule.

10. device as claimed in claim 8 is characterized in that, the structure of described mapping model comprises input layer, hidden layer and output layer, the transfer function of described hidden layer and output layer be Sigmoid type: f (x)=1/[1+e^ (x)];

Described acquisition module comprises:

Color value obtains submodule, is used to obtain N CMYK color value and corresponding RGB color value, and with its input and desired output initial data as learning sample;

The normalization submodule is used for normalized is carried out in described input and desired output initial data, makes its value between [0,1].

Images