CN106127702A - A kind of image mist elimination algorithm based on degree of depth study - Google Patents

Abstract

The invention discloses a kind of image mist elimination algorithm based on degree of depth study, have the fog in mist image to disturb for removing, reduce the fog impact on picture quality.Including: step 1, obtain training sample set and test sample collection；Step 2, is concentrated with mist image and carries out HSL spatial variations, extract and have the low bright feature of mist image local, and all characteristic components are carried out scaling and normalized sample；Step 3, finds out differentiation perspective rate, makes depth discrimination neural fusion confrontation type train；Step 4, utilizes the degree of depth to generate antagonism neutral net and is trained features described above component, and study is set up the mapping network between mist image and perspective rate；Step 5, uses the degree of depth to generate neutral net and test sample collection is carried out mist elimination test.Heretofore described mist elimination algorithm, sets up out mist image to the mapping relations between perspective rate by degree of depth learning algorithm, effectively solves the problem that conventional mist elimination algorithm prior information is not enough, have and preferably remove fog effect.

Description

A kind of image mist elimination algorithm based on degree of depth study

Technical field

The present invention relates to image processing techniques, pattern recognition and artificial intelligence field, be specifically related to a kind of based on the degree of depth The image mist elimination algorithm practised.

Background technology

In recent years, along with the development of artificial intelligence, increasing image capture device has under different scenes Automatically the ability of target is identified.But, the environmental factors restriction that this automatic identification ability is frequently subjected in scene, particularly mist Or the existence of haze, reduce saturation and the contrast of target in image, it has also become the automatic recognition effect of the equipment that affects important Factor, it is proposed that a kind of effective and feasible mist elimination algorithm has important theoretical and practical significance.

Traditional mist elimination algorithm is broadly divided into two classes: mist elimination algorithm based on atmosphere light scattering model image restoration with based on The mist elimination algorithm that image enhaucament is theoretical.The mist elimination algorithm of main flow is all based on what atmosphere light scattering model was set up at present, Qi Zhongying It is dark channel prior mist elimination algorithm with widest image defogging method.For the defogging method that presently, there are, have owing to lacking Effective priori in mist image, so all cannot obtain the perspective rate of optimum, presents color inclined during image restoration The phenomenon moved.Therefore, the perspective rate how obtaining optimum becomes the key issue in image mist elimination algorithm.

Summary of the invention

The present invention proposes a kind of image mist elimination algorithm based on degree of depth study, its object is to generate antagonism by the degree of depth Training sample is concentrated data to be trained study by neutral net, obtains there is the mapping network between mist characteristics of image and perspective rate, Find the perspective rate of adaptive optimal, carry out restoring conversion according to atmosphere light scattering model afterwards, finally realize image mist elimination.

It is an object of the invention to be realized by following technical proposals.

A kind of image mist elimination algorithm based on degree of depth study, comprises the following steps:

Step 1, sets up sample graph image set, including training sample set and test sample collection；Wherein for training sample set, adopt Collection fog there may be the image without mist under scene, and is manually atomized without mist image, is manually had mist image, finally With without mist image construction training sample set；For test sample collection, use true fog scene image；

Step 2, concentrates training sample and has mist image to carry out color notation conversion space, from RGB (red, green, blue) spatial alternation To HSL (hue, saturation, intensity) space, obtain the colourity of image, saturation and monochrome information characteristic component respectively, afterwards Under the rgb space of original image, obtain the low bright values in local and air light value, and all data are done at scaling and normalization Reason；

Step 3, calculates optimum perspective rate by training sample set, and in making it learn as the degree of depth, the degree of depth generates antagonism nerve The differentiation component of network training input, completes the training of network confrontation type jointly with another input quantity structure perspective rate；

Step 4, concentrates obtain to have mist image rgb space and HSL space characteristics component, and local based on training sample Low bright characteristic component and differentiation perspective rate, use the degree of depth in degree of deep learning algorithm to generate antagonism neutral net and be trained, To the perspective rate having in mist image, study is set up the mapping network between mist image and perspective rate.Wherein, the degree of depth generates antagonism god Generated neutral net through network by depth discrimination neutral net and the degree of depth to form；

Step 5, uses the above-mentioned degree of depth to generate neutral net to truly there being mist image to carry out mist elimination process.

Preferably, step 1 farther includes, and chooses image without mist that fog there may be under scene as initial training sample This collection, uses that repaiies that figure software obtains initial training sample to be manually atomized image, and same scene has mist image with without mist Image is as training sample set, and gathering afterwards has mist image as test sample collection in true fog scene.

Preferably, step 2 farther includes, and original sample is carried out eigentransformation, comprises the steps:

1) concentrate at training image, artificial atomization image is carried out HSL color notation conversion space, finally gives RGB Yu HSL two Plant the characteristic component in space；

2) the calculating process of the low bright values in the local of image block is:

$J_{low}\left(x\right)=\underset{x\∈\mathrm{\Ω}\left(x_0\right)}{min}\left(I^r\right(x),I^g(x),I^b(x\left)\right)---\left(1\right)$

In formula, r, g, b are respectively the triple channel of rgb space, Ω (x₀) represent with x₀Centered by regional area pixel Collection, J_lowX () is the low bright values of regional area；I^r(x)、I^g(x)、I^bX () indicates triple channel pixel in mist image rgb space respectively Value；

3) to above-mentioned RGB Yu HSL space characteristics component, and the low bright characteristic component in local carries out scaling so that it is dimension Number is 256*256, and is normalized each component, makes each component pixel point value belong to [0,1]；

4) air light value is:

A^c=min (max (I^c(x)),0.8) (2)

In formula, { r, g, b} are channel number in rgb space to c ∈, I^cX () indicates in mist image under c corresponding color passage Pixel value.

Preferably, step 3 farther includes, and differentiates that perspective rate is in described training:

$t_{dist}^c\left(x\right)=\left\{\begin{array}{cc}1,& J^c>I^c,J^c=I^c=A^c\\ \frac{A^c-I^c\left(x\right)}{A^c-J^c\left(x\right)},& J^c\&le;I^c<A^c\\ 0,& J^c\&NotEqual;I^c=A^c\end{array}\right.---\left(3\right)$

In formula, J^cX () is without the pixel value under c corresponding color passage in mist image.

Preferably, step 4 farther includes, and carries out the degree of depth and generates antagonism neural metwork training, comprises the steps:

1) construction depth differentiates neutral net, including 1 input layer, 5 convolutional layers, 4 batches of regularization layers, 1 entirely connect Connecing layer and 1 output layer, connected mode is: input layer → convolutional layer 1 → convolutional layer 2 → batch regularization layer 1 → convolutional layer 3 → batch Regularization layer 2 → convolutional layer 4 → batch regularization layer 3 → convolutional layer, 5 → batch regularization layer, 4 → full articulamentum → output layer, wherein, In depth discrimination neutral net, in convolutional layer to batch regularization layer conversion process, interlayer structure is not changed in；

Wherein, input layer is by structure perspective rateWith differentiation perspective rate t_distX () is respectively fed to depth discrimination nerve net In network, after multilamellar process of convolution, obtain characterizing the high-order feature of input information attribute, deliver to defeated after full articulamentum processes Going out layer, carry out discriminant classification by activation primitive Sigmoid function, wherein Sigmoid function is:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(4\right)$

2) construction depth generates neutral net, including 1 input layer, 2 convolutional layers, 2 warp laminations, 4 batches of canonicals Changing layer and 1 output layer, connected mode is: input layer → convolutional layer 1 → batch regularization layer 1 → convolutional layer, 2 → batch regularization layer 2 → warp lamination 1 → batch regularization layer 3 → warp lamination, 2 → batch regularization layer 4 → output layer, and output layer is for pond Layer, is the result exported behind batch regularization layer 4 pond.RGB, HSL are sent totally together with the low bright 7 layers of characteristic component in local by input layer Enter the degree of depth and generate in neutral net, reduce with mapping through hidden layer feature extraction, final output layer output construction perspective rate

3) in the training degree of depth generates antagonism nerve net, for ensureing the dependency between component, prior-constrained model is added, if The cost function set the goal:

$\begin{array}{cc}\underset{G}{min}& \underset{D}{\max }T(G,D)=\frac{1}{m}\underset{i=1}{\overset{m}{\&Sigma;}}\&lsqb;\log D\left(t_{dist}^{\left(i\right)}\right)+log(1-D(G\left(z^{\left(i\right)}\right)\left)\right)\&rsqb;\end{array}---\left(5\right)$

In formula, m represents the number of samples of minimum batch in training,Result is exported for depth discrimination neutral net,Represent the differentiation perspective rate of i-th sample, G (z⁽ⁱ⁾) it is that the degree of depth generates neutral net output result, z⁽ⁱ⁾Represent that the degree of depth generates I-th group of input sample characteristics in neutral net；

Generating antagonism neural metwork training in the degree of depth and include forward-propagating and two processes of back propagation, wherein forward passes Broadcast middle employing layering greedy algorithm to be trained, by training sample successively abstract, complete the extraction process of feature；Back propagation Middle employing SGD (stochastic gradient descent) parameter training algorithm, utilizes calibration information to carry out having monitor mode to learn, and generates the degree of depth Antagonism neutral net is carried out from pushing up to the regulation of end parameter.

Preferably, step 4.3 farther includes, and in training, definition differentiates label, and { 0,1}, i.e. calibration information is only with 0 or 1 Representing predetermined and differentiate result, wherein 1 represents that input data are to differentiate perspective rate, and 0 represents that input data are for structure perspective rate.For The degree of depth generates neutral net, after initialization, low to above-mentioned RGB, HSL and local bright 7 layers of characteristic component is sent into the degree of depth and generates nerve In network, output construction perspective rate is as the input quantity of depth discrimination neutral net, in the degree of depth generates neural metwork training, The calibration information of structure perspective rate is set to 1, in conjunction with differentiating result, utilizes SGD algorithm to carry out parameter adjustment, and wherein the degree of depth generates god In network, the gradient of undated parameter isFor depth discrimination neutral net, after initialization, Respectively differentiating that perspective rate inputs in depth discrimination neutral net with structure perspective rate, and differentiate that perspective rate calibration information is set to 1, structure perspective rate calibration information is set to 0, in conjunction with differentiating result, utilizes SGD algorithm to carry out parameter adjustment, wherein depth discrimination god In network, the gradient of undated parameter isBy depth discrimination nerve net Network generates neutral net and the calibration information of structure perspective rate carries out the different demarcation of 0 and 1 from the degree of depth so that the degree of depth generates nerve Network can obtain the perspective rate of optimum, and depth discrimination neutral net carries out Accurate classification to perspective rate, it is achieved it is right that the degree of depth generates The confrontation type training of anti-neutral net.

Preferably, step 5 farther includes, and obtains HSL space characteristics component, and extract test image from test image The low bright feature in local, afterwards RGB, HSL with local low bright characteristic component send into above-mentioned training the degree of depth generate neutral net In, obtain constructing perspective rateCarry out mist elimination conversion according to the following formula:

${\tilde{J}}^c\left(x\right)=\left\{\begin{array}{cc}\frac{I^c-(1-{\tilde{t}}^c)A^c}{{\tilde{t}}^c},& {\tilde{t}}^c\&NotEqual;0\\ \frac{I^c-0.92\&times;A^c}{0.08}& {\tilde{t}}^c=0\end{array}\right.---\left(6\right)$

In formula,It is image after mist elimination corresponding for c for channel layer,The output of neutral net is generated for the c correspondence degree of depth Perspective rate.

The present invention combines the low bright feature in local, it is proposed that a kind of image mist elimination algorithm based on degree of depth study.Calculate with tradition Method is compared, institute of the present invention extracting method, it is not necessary to have substantial amounts of prior information in mist image, learn by substantial amounts of training sample There is the perspective rate in mist image, be finally completed mist image to changing without mist image accordingly.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the present invention；

Fig. 2 is that the degree of depth of the present invention generates antagonism neural network training process schematic diagram；

Fig. 3 is depth discrimination neural network structure figure of the present invention；

Fig. 4 is that the degree of depth of the present invention generates neural network structure figure.

Detailed description of the invention

Below in conjunction with the accompanying drawings and detailed description of the invention further illustrates the present invention, but the embodiment that this kind is described by accompanying drawing It is exemplary, is only used for explaining the present invention, it is impossible to limit the scope of the claims in the present invention protection.

The present invention a kind of based on the degree of depth study image mist elimination algorithm as it is shown in figure 1, in figure solid line represent forward-propagating Journey, dotted line represents back-propagation process, and its key step is described below:

1, training sample set and test sample collection are obtained

Choosing fog and there may be under environment, indoor, outdoor scene image without mist, as initial sample set, uses Initial sample set is manually atomized by Photoshop software, obtains there is mist image, with initial sample under corresponding scene and illumination This collection composition training sample set.Gathering afterwards has mist image as test sample collection in true fog environment.

2, the degree of depth generates neutral net input feature vector component extraction

Carrying out eigentransformation by initial sample, each input feature vector extracted in training degree of depth generation neutral net divides Amount, comprises the steps:

1) concentrate at training image, artificial atomization image is carried out HSL color notation conversion space, finally gives RGB Yu HSL two Plant the characteristic component in space；

2) combine the low bright characteristic in local, for training network adds the low bright feature in local, extract the low bright characteristic component in local Calculating process be:

$J_{low}\left(x\right)=\underset{x\&Element;\mathrm{\&Omega;}\left(x_0\right)}{min}\left(I^r\right(x),I^g(x),I^b(x\left)\right)---\left(1\right)$

In formula, r, g, b are respectively the triple channel of rgb space, Ω (x₀) represent with x₀Centered by regional area pixel Collection, J_lowX () is the low bright values of regional area；I^r(x)、I^g(x)、I^bX () indicates in mist image triple channel picture in rgb space respectively Element value；

3) to above-mentioned RGB Yu HSL space characteristics component, and the low bright characteristic component in local carries out scaling so that it is dimension Number is 256*256, and each component carries out linear normalization process, makes each component pixel point value belong to [0,1]；

4) air light value is selected

In training sample, calculate the maximum having each passage of mist image, as the air light value A of this passage, for avoiding Air light value is affected by local highlight regions in image, sets its upper limit threshold as 0.8, and in single channel, air light value is selected concrete Formula is:

A^c=min (max (I^c(x)),0.8) (2)

In formula, { r, g, b} are channel number in rgb space to c ∈, I^cX () indicates in mist image under c corresponding color passage Pixel value.

3, depth discrimination neutral net input feature vector component extraction

Optimum perspective rate is calculated so that it is in learning as the degree of depth, the degree of depth generates antagonism neutral net instruction by training sample set Practice the differentiation component of input, jointly complete the training of network confrontation type with another input quantity structure perspective rate.Make training sample set In have mist image and without between mist image change perspective rate as differentiate perspective rate, in described training differentiate perspective rate be:

$t_{dist}^c\left(x\right)=\left\{\begin{array}{cc}1,& J^c>I^c,J^c=I^c=A^c\\ \frac{A^c-I^c\left(x\right)}{A^c-J^c\left(x\right)},& J^c\&le;I^c<A^c\\ 0,& J^c\&NotEqual;I^c=A^c\end{array}\right.---\left(3\right)$

In formula, { r, g, b} are channel number in rgb space to c ∈, I^cX () is for there to be Color Channel corresponding for c in mist image.

4, the degree of depth generates antagonism neural metwork training

Concentrate obtain to have mist image rgb space and HSL space characteristics component, and the low bright spy in local based on training sample Levy component and differentiate perspective rate, using the degree of depth in degree of deep learning algorithm to generate antagonism neutral net and be trained, obtain there is mist Perspective rate in image, study is set up the mapping network between mist image and perspective rate.Wherein, the degree of depth generates antagonism neutral net Generated neutral net by depth discrimination neutral net and the degree of depth to form.The degree of depth generates antagonism neural network training process such as Fig. 2 institute Show, wherein F_distRepresent the extraction process differentiating perspective rate, realize according to formula (3), F_dedazeRepresent and completed by perspective rate Mist image, to the conversion without mist image, realizes according to formula (6), G Yu D represents that the degree of depth generates neutral net and depth discrimination respectively Neutral net, implementing step is:

1) construction depth differentiate neutral net, each interlayer connected mode as shown in Figure 3: input layer → convolutional layer 1 → convolution Layer 2 → batch regularization layer 1 → convolutional layer, 3 → batch regularization layer 2 → convolutional layer, 4 → batch regularization layer 3 → convolutional layer, 5 → batch canonical Change layer 4 → full articulamentum → output layer.It is worthy of note: due in the operation of convolutional layer to batch regularization layer, interlayer structure Be not changed in, thus in Fig. 3 respectively convolutional layer 2 and batch regularization layer 1, convolutional layer 3 and batch regularization layer 2, convolutional layer 4 with batch just Then change layer 3, convolutional layer 5 is placed in same structure sheaf with batch regularization layer 4.

Wherein, input layer data are three layers of characteristic component, for the three-channel characteristic component of RGB, structure perspective rate With differentiation perspective rate t_distX () is respectively fed in depth discrimination neutral net as input, by convolution operation.First order convolution Select the convolution kernel of 8*8*48；Second level convolution selects the convolution kernel of 6*6*48, processes through batch regular sheaf, then after convolution output Carrying out ReLU activation primitive operation, this operation is all passed through in the output of follow-up convolution；Third level convolutional layer selects the convolution kernel of 4*4*64； Fourth stage convolution selects the convolution kernel of 5*5*80；Level V convolution selects the convolution kernel of 7*7*80；After 5 convolution operation, defeated Go out and carry out full attended operation, send into output layer afterwards, carry out discriminant classification by activation primitive Sigmoid function, wherein Sigmoid function is:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(4\right)$

2) construction depth generates neutral net, each interlayer connected mode as shown in Figure 4: input layer → convolutional layer 1 → batch just Then change layer 1 → convolutional layer 2 → batch regularization layer 2 → warp lamination, 1 → batch regularization layer 3 → warp lamination, 2 → batch regularization layer 4 → output layer.It is worthy of note: owing to, in convolutional layer (or warp lamination) to the operation of batch regularization layer, interlayer structure does not has Change, therefore Fig. 4 is respectively convolutional layer 1 and batch regularization layer 1, convolutional layer 2 and batch regularization layer 2, warp lamination 1 and batch regularization Layer 3, warp lamination 2 and batch regularization layer 4 are placed in same structure sheaf.Wherein, input layer be input sample data RGB, HSL and 7 layers of characteristic component of the lowest bright feature, first order convolution selects the convolution kernel of 9*9*64；The convolution of second level convolution 4*4*128 Core, carries out deconvolution operation afterwards.The convolution kernel of 9*9*224 is selected in first order deconvolution, and step-length is 3；Second level deconvolution choosing With the convolution kernel of 11*11*3, step-length is 2.The most all convolutional layers all process through batch regular sheaf, then after exporting with warp lamination Carry out the operation of ReLU (rectification linear unit) activation primitive.Reduce, to defeated with the mapping of deconvolution through the feature extraction of convolution Going out data and carry out the process of average pondization, select 18*18 template that data are carried out pond, step-length is 1, and in the present invention, pondization operation is main In network to be solved, image generates the grid problem occurred, and structure perspective rate pondization exportedAs output layer；

3) for ensureing the dependency between component, in the training degree of depth generates antagonism nerve net, prior-constrained model is added, if The cost function set the goal:

$\begin{array}{cc}\underset{G}{min}& \underset{D}{\max }T(G,D)=\frac{1}{m}\underset{i=1}{\overset{m}{\&Sigma;}}\&lsqb;\log D\left(t_{dist}^{\left(i\right)}\right)+log(1-D(G\left(z^{\left(i\right)}\right)\left)\right)\&rsqb;\end{array}---\left(5\right)$

In formula, m represents the number of samples of minimum batch in training,Result is exported for depth discrimination neutral net,Represent the differentiation perspective rate of i-th sample, G (z⁽ⁱ⁾) it is that the degree of depth generates neutral net output result, z⁽ⁱ⁾Represent that the degree of depth generates I-th group of input sample characteristics in neutral net；

4) generate antagonism neural metwork training in the degree of depth and include that the degree of depth generates neutral net and depth discrimination neutral net Training, first initialize entire depth generate antagonism neural network parameter, parameter initial rules is: random parameter in convolutional layer Meeting average is 0, and variance is the normal distribution of 0.02, and in batch regularization layer, random parameter meets average is 1, and variance is 0.02 just State is distributed.Set afterwards and differentiate that { 0,1}, wherein 1 represents that input data are to differentiate perspective rate to label, and 0 represents that input data are structure Make perspective rate.The degree of depth generates in neural metwork training, above-mentioned 7 layers of characteristic component is sent into the degree of depth and generates in neutral net, output Structure perspective rate is as the input quantity of depth discrimination neutral net, in this degree of depth generates neutral net, and the mark of structure perspective rate Determining information and be set to 1, in conjunction with differentiating result, utilize SGD algorithm to carry out parameter adjustment, wherein the gradient of undated parameter isIn depth discrimination neutral net, respectively differentiating perspective rate and structure perspective rate input depth discrimination In neutral net, and differentiating that perspective rate calibration information is set to 1, structure perspective rate calibration information is set to 0, in conjunction with differentiating result, Utilizing SGD algorithm to carry out parameter adjustment, wherein the gradient of undated parameter isLogical Cross depth discrimination neutral net and the degree of depth to generate neutral net the calibration information of structure perspective rate carries out the different demarcation of 0 and 1, Making the degree of depth generate neutral net and can obtain the perspective rate of optimum, perspective rate is accurately divided by depth discrimination neutral net Class, it is achieved the degree of depth generates the confrontation type training of antagonism neutral net.

5, image mist elimination test

There is mist image to carry out HSL conversion test sample concentration, obtain the characteristic component under HSL space, and extract survey Attempt the low bright feature in local of picture, afterwards refreshing with the degree of depth generation that the low bright characteristic component in local sends into above-mentioned training to RGB, HSL In network, obtain constructing perspective rateFollow-up carry out mist elimination process according to atmosphere light value model, carry out mist elimination according to the following formula Conversion:

${\tilde{J}}^c\left(x\right)=\left\{\begin{array}{cc}\frac{I^c-(1-{\tilde{t}}^c)A^c}{{\tilde{t}}^c},& {\tilde{t}}^c\&NotEqual;0\\ \frac{I^c-0.92\&times;A^c}{0.08}& {\tilde{t}}^c=0\end{array}\right.---\left(6\right)$

In formula,It is image after mist elimination corresponding for c for channel layer,The output of neutral net is generated for the c correspondence degree of depth Perspective rate.

Above-described embodiment is the specific embodiment of the present invention, and institute is it will be clear that embodiments of the present invention not only limiting It is formed on the replacement of above-described embodiment, the follow-up various changes carried out in embodiment and similarity method, all without departing from Scope defined in appended claims.

Claims (7)

1. an image mist elimination algorithm based on degree of depth study, it is characterised in that comprise the following steps:

Step 1, sets up sample graph image set, including training sample set and test sample collection；

Wherein, for training sample set, gather fog and there may be the image without mist under scene, and carry out manually without mist image Atomization, is manually had mist image, final with without mist image construction training sample set；

For test sample collection, use true fog scene image；

Step 2, concentrates training sample and has mist image to carry out color notation conversion space, transform to HSL space from rgb space, respectively Obtain the colourity of image, saturation and monochrome information characteristic component, under the rgb space of original image, obtain the low bright values in local afterwards With air light value, and all data are done scaling and normalized；

Step 3, calculates optimum perspective rate by training sample set, and in making it learn as the degree of depth, the degree of depth generates antagonism neutral net The differentiation component of training input, completes the training of network confrontation type jointly with another input quantity structure perspective rate；

Step 4, concentrates based on training sample obtain to have mist image rgb space and HSL space characteristics component, and local is low bright Characteristic component and differentiation perspective rate, use the degree of depth in degree of deep learning algorithm to generate antagonism neutral net and be trained, had Perspective rate in mist image, study is set up the mapping network between mist image and perspective rate；Wherein, the degree of depth generates antagonism nerve net Network is generated neutral net by depth discrimination neutral net and the degree of depth and forms；

Step 5, uses the above-mentioned degree of depth to generate neutral net to truly there being mist image to carry out mist elimination process.

Method the most according to claim 1, it is characterised in that: in step 1, choose fog there may be under scene without mist Image as initial training sample set, uses that repaiies that figure software obtains initial training sample to be manually atomized image, and identical field Scape has mist image with without mist image as training sample set, gathering afterwards has mist image as test in true fog scene Sample set.

Method the most according to claim 1, it is characterised in that:

Step 2 farther includes sub-step:

2.1 concentrate at training image, artificial atomization image is carried out HSL color notation conversion space, finally gives RGB Yu HSL two kinds Characteristic component in space；

The calculating process of the low bright values in local of 2.2 image blocks is:

$J_{low}\left(x\right)=\underset{x\&Element;\mathrm{\&Omega;}\left(x_0\right)}{min}\left(I^r\right(x),I^g(x),I^b(x\left)\right)---\left(1\right)$

In formula, r, g, b are respectively the triple channel of rgb space, Ω (x₀) represent with x₀Centered by regional area pixel point set, J_low X () is the low bright values of regional area；I^r(x)、I^g(x)、I^bX () indicates triple channel pixel value in mist image rgb space respectively；

2.3 pairs of above-mentioned RGB Yu HSL space characteristics components, and locally low bright characteristic component carries out scaling so that it is dimension For 256*256, and each component is normalized, makes each component pixel point value belong to [0,1]；

2.4 air light values are:

A^c=min (max (I^c(x)),0.8) (2)

In formula, { r, g, b} are channel number in rgb space to c ∈, I^cX () indicates the picture in mist image under c corresponding color passage Element value.

Method the most according to claim 1, it is characterised in that: in step 3, described differentiation perspective rate is obtained by following formula:

$t_{dist}^c\left(x\right)=\left\{\begin{array}{cc}1,& J^c>I^c,J^c=I^c=A^c\\ \frac{A^c-I^c\left(x\right)}{A^c-J^c\left(x\right)},& J^c\&le;I^c<A^c\\ 0,& J^c\&NotEqual;I^c=A^c\end{array}\right.---\left(3\right)$

In formula, J^cX () is without the pixel value under c corresponding color passage in mist image.

Method the most according to claim 1, it is characterised in that:

Step 4 farther includes sub-step:

4.1 construction depths differentiate neutral net, including 1 input layer, 5 convolutional layers, 4 batches of regularization layers, 1 full connection Layer and 1 output layer, connected mode is: input layer → convolutional layer 1 → convolutional layer 2 → batch regularization layer 1 → convolutional layer 3 → batch just Then change layer 2 → convolutional layer 4 → batch regularization layer 3 → convolutional layer, 5 → batch regularization layer, 4 → full articulamentum → output layer；

Wherein, input layer is by structure perspective rate t (x) and differentiation perspective rate t_distX () is respectively fed in depth discrimination neutral net, After multilamellar process of convolution, obtain characterizing the high-order feature of input information attribute, after full articulamentum processes, deliver to output layer, Carrying out discriminant classification by activation primitive Sigmoid function, wherein Sigmoid function is:

$f\left(x\right)=\frac{1}{1+e^{-x}}---\left(4\right)$

4.2 construction depths generate neutral net, including 1 input layer, 2 convolutional layers, 2 warp laminations, 4 batches of regularizations Layer and 1 output layer, connected mode is: input layer → convolutional layer 1 → batch regularization layer 1 → convolutional layer 2 → batch regularization layer 2 → Warp lamination 1 → batch regularization layer 3 → warp lamination, 2 → batch regularization layer 4 → output layer, wherein, output layer is batch regularization The pond layer that layer 4 exports after carrying out pond；

Wherein, RGB, HSL are sent into the degree of depth totally together with the low bright 7 layers of characteristic component in local and generate in neutral net by input layer, warp Hidden layer feature extraction is reduced with mapping, final output layer output construction perspective rate

4.3 in the training degree of depth generates antagonism nerve net, for ensureing the dependency between component, adds prior-constrained model, sets The cost function of target:

$\underset{G}{min}\underset{G}{\max }T(G,D)=\frac{1}{m}\underset{i=1}{\overset{m}{\&Sigma;}}\&lsqb;\log D\left(t_{dist}^{\left(i\right)}\right)+log(1-D(G\left(z^{\left(i\right)}\right)\left)\right)\&rsqb;---\left(5\right)$

In formula, m represents the number of samples of minimum batch in training,Result is exported for depth discrimination neutral net,Table Show the differentiation perspective rate of i-th sample, G (z⁽ⁱ⁾) it is that the degree of depth generates neutral net output result, z⁽ⁱ⁾Represent that the degree of depth generates nerve I-th group of input sample characteristics in network；

Generate antagonism neural metwork training in the degree of depth and include forward-propagating and two processes of back propagation, wherein in forward-propagating Use layering greedy algorithm to be trained, by training sample successively abstract, complete the extraction process of feature；Back propagation is adopted With stochastic gradient descent parameter training algorithm, utilize calibration information to carry out having monitor mode to learn, the degree of depth is generated antagonism nerve Network is carried out from the parameter regulation pushed up the end of to.

Method the most according to claim 5, it is characterised in that: in step 4.3, in training, definition differentiation label 0,1}, I.e. calibration information only represents predetermined with 0 or 1 and differentiates result, and wherein 1 represents that input data are to differentiate perspective rate, and 0 represents input number According to for constructing perspective rate；

Neutral net is generated for the degree of depth, after initialization, RGB, HSL and local 7 layers of characteristic component of low amounts is sent into the degree of depth and generate In neutral net, output construction perspective rate is as the input quantity of depth discrimination neutral net；

In the degree of depth generates neural metwork training, the calibration information of structure perspective rate is set to 1, in conjunction with differentiating result, utilizes SGD Algorithm carries out parameter adjustment, and during wherein the degree of depth generates neutral net, the gradient of undated parameter is:

$\&dtri;\frac{1}{m}\underset{i=1}{\overset{m}{\&Sigma;}}log(1-D(G\left(z^{\left(i\right)}\right)\left)\right);$

For depth discrimination neutral net, after initialization, respectively differentiating perspective rate and structure perspective rate input depth discrimination god In network, and differentiating that perspective rate calibration information is set to 1, structure perspective rate calibration information is set to 0, in conjunction with differentiating result, and profit Parameter adjustment is carried out with SGD algorithm；

Wherein in depth discrimination neutral net, the gradient of undated parameter is:

$\&dtri;\frac{1}{m}\underset{i=1}{\overset{m}{\&Sigma;}}\&lsqb;logD\left(t_{dist}^{\left(i\right)}\right)+log(1-D(G\left(z^{\left(i\right)}\right)\left)\right)\&rsqb;$

Generate neutral net by depth discrimination neutral net and the degree of depth and the calibration information of structure perspective rate is carried out 0 and 1 not With demarcating so that the degree of depth generates neutral net can obtain the perspective rate of optimum, and perspective rate is carried out by depth discrimination neutral net Accurate classification, it is achieved the degree of depth generates the confrontation type training of antagonism neutral net.

Method the most according to claim 1, it is characterised in that: in step 5, from test image, obtain HSL space characteristics Component, and extract the low bright feature in local of test image, afterwards RGB, HSL are sent into above-mentioned training with the low bright characteristic component in local The degree of depth generate in neutral net, obtain constructing perspective rateCarry out mist elimination conversion according to the following formula:

${\tilde{J}}^c\left(x\right)=\left\{\begin{array}{cc}\frac{I^c-(1-{\tilde{t}}^c)A^c}{{\tilde{t}}^c},& {\tilde{t}}^c\&NotEqual;0\\ \frac{I^c-0.92\&times;A^c}{0.08}& {\tilde{t}}^c=0\end{array}\right.---\left(6\right)$

In formula,It is image after c correspondence mist elimination for channel number,The perspective of the output of neutral net is generated for the c correspondence degree of depth Rate.