CN106503734A - Based on trilateral filter and the image classification method of the sparse autocoder of storehouse - Google Patents

Abstract

The present invention is to provide a kind of based on trilateral filter and the image classification method of the sparse autocoder of storehouse.First, smooth image is obtained using trilateral filter, filter Gauss, speckle and the impulsive noise of degraded image while the spectral space feature of the pixel for extracting described image；Secondly, high-order feature extraction is carried out using improved storehouse sparse autocoder；Finally, carried out supervising trim network and classification using random forest grader.The present invention is incorporated into sparse for improved storehouse autocoder and random forest grader in Hyperspectral data classification, as a kind of deep learning framework, the sparse autocoder of improved storehouse can extract the abstract He useful profound feature of spectroscopic data layer by layer, so as to improve the classification performance of spectroscopic data.The present invention is applicable not only to classify high spectrum image, while can also classify to other images.There is very strong portability, be more easy to the demand for meeting image classification.

Description

Based on trilateral filter and the image classification method of the sparse autocoder of storehouse

Technical field

The present invention relates to a kind of image classification method, particularly a kind of joint trilateral filter and deep learning theory Hyperspectral image classification method.

Background technology

High-spectrum remote-sensing obtains view data using a lot of very narrow electromagnetic wave bands from object interested, typically from Light is provided with tens in the range of heat wave infrared band to hundreds of continuous wave bands, and its spectral resolution may be up to a nanometer quantity Level.For the pixel that each is recorded, abundant spectral information can provide the complete light profiling of observation earth's surface and spy Property, an effective instrument therefore can be counted as when material is distinguished.High light spectrum image-forming has become a kind of important at present Remotely sensed image detection means, its essence be can and meanwhile provide atural object distribution spatial information and high-resolution spectrum Information.Although hyperspectral image data has data more significant advantage and development potentiality than ever, to high spectrum image Data carry out, in spatial information and withdrawing spectral information and application process, occurring in that much technical difficult problems, to researchers Bring huge challenge.On the other hand, during being captured using bloom spectrum sensor and transmitting high spectrum image, generally high This noise and impulsive noise etc. can be incorporated in image, cause the image quality decrease for producing, so as to largely restrict figure The nicety of grading of picture.Therefore, how to effectively improve picture quality and extract the spectral-spatial information of pixel, so as to realize height Effective classification of spectroscopic data obtains more and more extensive concern.

In to classification hyperspectral imagery, remove picture noise and extraction spectral-spatial feature is particularly important.This is because Noise causes the quality serious degradation of image, so as to largely restrict the nicety of grading of image.Secondly, partial zones Pixel in domain typically represents identical material and with similar spectral characteristic, therefore comprehensively utilizes the space of high-spectral data Information is expected to obtain more preferable classification performance.However, existing document generally only make use of the spectral-spatial information of pixel, Do not account for the geometry propinquity of pixel in regional area, and the Gauss of degraded image, speckle and pulse can not be effectively filtered out Noise etc..

On the other hand, grader mostly more popular at present is often viewed as shallow-layer learning model, such as linear Hold vector machine, logistic regression, maximum likelihood classifier and rarefaction representation base grader etc..But these graders can not be effectively Extract the high-order feature of spectroscopic data.

Content of the invention

It is an object of the invention to provide a kind of classification performance that can improve spectroscopic data based on trilateral filter and The image classification method of the sparse autocoder of storehouse.

Invention is realized in：

First, smooth image is obtained using trilateral filter, extract the spectral-spatial feature of the pixel of described image Gauss, speckle and the impulsive noise of degraded image are filtered simultaneously；

Secondly, high-order feature extraction is carried out using the sparse autocoder of improved storehouse (SSA)；

Finally, carried out supervising trim network and classification using random forest grader.

The present invention can also include：

1st, described image is high spectrum image.

2nd, the pixel of the trilateral filter more new formula is：

Wherein, t and v represent the coordinate of pixel, S respectively_pRepresent the set of the neighborhood of pixel t, δ_s、δ_rAnd δ_dRespectively control Spatial parameter processed, magnitude parameters and pulse parameter, I (t) are the center pixels of initial pictures, and I (v) is the picture in neighborhood window Element, J (t) and J (j) are the pixel in the center pixel based on reference picture and neighborhood window respectively, W_tIt is that spacial proximity adds PowerSimilarity is weightedWeight with pulseProduct, W_tIt is further represented as：

When pulse weight factor is executed, using grade sequence absolute difference (Rank-Ordered Absolute Differences (ROAD)) determining that sample point is the marginal point of image or the pixel polluted by impulsive noise, accordingly Computing formula is as follows：

D [t, v]=| I (t)-I (v) |

r_f(t, v)=f^thsmallest d[t,v]

Wherein, r_f(t, v) represents antipode value d [t, v] of increasing arrangement, i.e., add up from small to large to choose f d [t, V] value, 2≤g≤w²- 2, w are window sizes, in equationIn, ROAD statistics is used in tolerance Imago element and the proximity of g-th neighborhood territory pixel.

3rd, carry out high-order feature extraction using the sparse autocoder of improved storehouse to specifically include：

Sparse autocoder is attempted extracting the input for hiding feature so that the reconstructed vector of decoding layer is similar to input layer Data, i.e. input vector x ∈ R^DIt is mapped to hidden layer and produces a linear composition z ∈ R^S, then non-linear sharp using one Function f (x)=(1+exp (- x)) living^-1Output a is obtained, as follows

Z=wx+b₁, a=f (z)

Wherein, w ∈ R^S×DAnd b₁∈R^S×1Input layer is represented respectively to weight and the biasing of hidden layer, and D represents input data Dimension, S represent hidden layer neuron number, in order to verify whether that the reconstruct y for obtaining is similar to initial input, a for obtaining are entered Row decoding, as follows：

T=va+b₂, y=f (t)

Wherein, v ∈ R^D×S、b₂∈R^D×1Hidden layer is represented respectively to weight and the biasing of output layer, initial input x and reconstruct Minimum error between yCalculated；In order to find some ad hoc structures and the phase of initial input Guan Xing, sparse constraint and weight penalty term are added in network, and object function J is as follows：

Wherein, M is number of samples, x_mAnd y_mM-th input and output are represented respectively, and λ representation severely punishes penalty parameter, and η is represented The weight of sparse penalty term,Company between j-th input neuron of i-th hidden neuron and l-1 layers that represent l layers Connect weight；Additionally,Represent the average activation primitive r of target of hidden unit Average activation primitive with hidden neuron i

Using the object function of the sparse autocoder of each layer of improved artificial fish school intelligent optimizing algorithm optimization, specifically Scheme is as follows：

If shoal of fish sum is n, each fish represents the difference and additivity value of a neutral net and any two fishes still A neutral net is represented, the parameters optimization of sparse autocoder includes two weight matrix [w] and [v] and threshold vector b； In current position, the food concentration of individual fish is Y=1/E, two individual fish X_pAnd X_qThe distance between as follows：

Wherein, output layer neuron number is N, w_ij(p) and w_ijQ () represents Artificial Fish X respectively_pAnd X_qParameter matrix I-th row, the element of jth row, v in [w]_ki(p) and v_kiQ () represents Artificial Fish X respectively_pAnd X_qParameter matrix [v] in row k, The element of the i-th row, b_i(p) and b_iQ () represents Artificial Fish X respectively_pAnd X_qThreshold vector b in the i-th row element, b_k(p) and b_k Q () represents Artificial Fish X respectively_pAnd X_qThreshold vector b in row k element, d_pqRepresent two individual fish X_pAnd X_qBetween away from From.

Foraging behavior：W (p) is the current state of individual fish, then being randomly chosen new state w (q) within sweep of the eye, If food concentration Y_q＞ Y_p, w (p) is according to formula It is updated；Otherwise, select a new state again to judge whether to meet advance condition, if executing maximum trial time Still advance condition is unsatisfactory for after number, and individual fish will be according to formula w (pnext)=w (p)+rand () × step+ χ × (w_ij (best)-w_ij(l)),if(Y_q≤Y_p) random behavior is executed,

Wherein, w (pnext) is the next step state of individual fish, and rand () is equably to produce one between [0,1] at random Number, w_ijAnd w (best)_ijL () represents current optimum Artificial Fish X respectively_bestWith random fish X_lParameter matrix [w] in the i-th row, The element of j row, χ is golden ratio coefficient lambda and skewProduct representation be：

Wherein,Represent the equally distributed random value of standard and interval is

The visual field is dynamically adjusted using fuzzy logic system apart from V and moving step length M, the dispersion in fish school location and repeatedly Generation number is used as input variable, while V and M is used as output variable, dispersion measure expression formula is as follows：

Wherein, shoal of fish sum is n, and k represents current iteration,WithArtificial Fish X during kth time iteration is represented respectively_p With optimum Artificial Fish X_best, D represents dispersibility of the input variable of fuzzy system i.e. between the shoal of fish, iteration and dispersibility tolerance quilt Normalization is between 0 to 1, as follows：

K_Norm=k/K_max

In formula, K_NormAnd D_NormCurrent iteration number of times k and dispersibility D is carried out the fuzzy system after standard normal respectively System |input paramete；K_maxFor maximum iteration time；D_minAnd D_maxThe minima of dispersibility and maximum in respectively each iteration.

Sparse for improved storehouse autocoder and random forest grader are incorporated into Hyperspectral data classification by the present invention In, used as a kind of deep learning framework, the sparse autocoder of improved storehouse can extract the abstract of spectroscopic data layer by layer And useful profound feature, so as to improve the classification performance of spectroscopic data.

It is right that joint trilateral filter proposed by the present invention is applicable not only to the sorting technique of the sparse autocoder of storehouse High spectrum image is classified, while can also classify to other images.There is very strong portability, be more easy to satisfaction figure Demand as classification.

Advantages of the present invention is as follows：

1) similarity weight w is being calculated_rDuring, use the reference picture based on dual-tree complex wavelet transform, and not It is initial high spectrum image.The stability that the realization of this step can improve gray scale Likelihood Computation is more accurate so as to obtain As a result.

2) in pixel more new formula, pulse weight factor is introduced.By considering the spectral similarity of pixel and several What propinquity, so as to effectively obtain the spectrum and space characteristics of high spectrum image, while filter image to greatest extent producing The Gaussian noise produced in raw and transmitting procedure and impulsive noise etc..

3) character representation of smooth image is successively extracted using improved sparse autocoder (ISA), using Artificial Fish Colony intelligence optimized algorithm optimizes the object function of ISA, and the back-propagation algorithm so as to reduce traditional is easy to be absorbed in local extremum Probability accelerates convergence of algorithm speed simultaneously.Spectroscopic data is adaptively extracted using the sparse autocoder of storehouse further High-order feature.Possibility is provided for obtaining optimal classification accuracy.

Description of the drawings

Fig. 1 is the classification hyperspectral imagery flow chart of a width spectral-spatial information consolidation；

Fig. 2 is the flow chart of the sparse autocoder of a width artificial fish school algorithm；

Fig. 3 is Artificial Fish model；

Fig. 4 (a)-Fig. 4 (c) is the input of fuzzy logic system and output variable；

Fig. 5 is that the sky based on trilateral filter and the sparse autocoder of storehouse composes united classification hyperspectral imagery frame Frame；

Fig. 6 (a) is the Real profiles of nine kinds of atural objects of Indiana farm data；

Fig. 6 (b) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm OS-SVM；

Fig. 6 (c) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm BF-SVM；

Fig. 6 (d) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm TF-SVM；

Fig. 6 (e) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm OS-SSAR；

Fig. 6 (f) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm BF-SSARF；

Fig. 6 (g) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm TF-SSARF；

Fig. 6 (h) is the overall classification accuracy figure of the Indiana farm data based on contrast algorithm TF-ISSARF；

Fig. 7 (a) is the Real profiles of 13 kinds of atural objects of Kennedy Space Center's data；

Fig. 7 (b) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm OS-SVM；

Fig. 7 (c) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm BF-SVM；

Fig. 7 (d) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm TF-SVM；

Fig. 7 (e) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm OS-SSARF；

Fig. 7 (f) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm BF-SSARF；

Fig. 7 (g) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm TF-SSARF；

Fig. 7 (h) is the overall classification accuracy figure of the Kennedy Space Center's data based on contrast algorithm TF-ISSARF；

Fig. 8 (a) describes the spatial parameter based on trilateral filter, and magnitude parameters and pulse parameter are to Indiana farm number According to overall classification accuracy impact；

Fig. 8 (b) describes the spatial parameter based on trilateral filter, and magnitude parameters and pulse parameter are to Kennedy Space Center The impact of the overall classification accuracy of data；

Fig. 9 (a) describes the window size of the spacial proximity weighting based on trilateral filter to Indiana farm data Overall classification accuracy impact；

Fig. 9 (b) describes the window size of the spacial proximity weighting based on trilateral filter to Kennedy Space Center's number According to overall classification accuracy impact.

Specific embodiment

Illustrate below and the present invention is described in more detail.

First, before classification hyperspectral imagery is executed, smooth high spectrum image is obtained using trilateral filter.Extract Gauss, speckle and impulsive noise of degraded image etc. are filtered while the spectral-spatial feature of pixel.

Secondly, high-order feature extraction is carried out using the sparse autocoder of improved storehouse (SSA).

Finally, carried out supervising trim network and classification using random forest grader.

It is right that joint trilateral filter proposed by the present invention is applicable not only to the sorting technique of the sparse autocoder of storehouse High spectrum image is classified, while can also classify to other images.There is very strong portability, be more easy to satisfaction figure Demand as classification.

The present invention is specifically included：

1st, the present invention proposes to carry out smothing filtering with trilateral filter to initial degraded image.The three sides filtering for proposing The pixel of device more new formula is as follows：

Wherein, t and v represent the coordinate of pixel respectively.S_pRepresent the set of the neighborhood of pixel t.δ_s,δ_rAnd δ_dControl respectively Spatial parameter, magnitude parameters and pulse parameter.It is pixel in neighborhood window that I (t) is the center pixel and I (v) of initial pictures. J (t) and J (j) are the pixel in center pixel and neighborhood window based on reference picture respectively.W_tIt is spacial proximity weightingSimilarity is weightedWeight with pulseProduct, can be further represented as：

When pulse weight factor is executed, using grade sequence absolute difference (Rank-Ordered Absolute Differences (ROAD)) determining that sample point is the marginal point of image or the pixel polluted by impulsive noise.Accordingly Computing formula is as follows：

D [t, v]=| I (t)-I (v) | (3)

r_f(t, v)=f^thsmallest d[t,v] (4)

Wherein, r_f(t, v) represents antipode value d [t, v] of increasing arrangement, i.e., add up from small to large to choose f d [t, V] value.2≤g≤w²- 2, w are window sizes.In equation (5), ROAD statistics is used to measure center pixel and g-th neighborhood The proximity of pixel.

Generally, if there is no noise pixel in image, in neighborhood window, most of pixel represents similar intensity Value and produce low ROAD values.If conversely, pixel is polluted by impulsive noise, being contaminated spectral intensity and the neighborhood territory pixel of pixel There is larger difference, cause to produce high ROAD functional values.Especially, if at least exist in image a noise sample or High ROAD values are produced, now pulse weight is activated and passes through formula (1) and updates current pixel value.If on the contrary, figure There is no pulse pixel or high ROAD values as in, pulse weight factor is suppressed, i.e. w_d=0, now using bilateral filtering Device carries out smothing filtering to initial pictures.

2nd, obtained after smooth high spectrum image using above-mentioned trilateral filter, propose a kind of improved storehouse sparse automatically Encoder adaptively extracts the high-order character representation of smoothed image.

Shown in being implemented as follows：

Sparse autocoder is attempted extracting the input for hiding feature so that the reconstructed vector of decoding layer is similar to input layer Data.That is input vector x ∈ R^DIt is mapped to hidden layer and produces a linear composition z ∈ R^S, then non-linear sharp using one Function f (x)=(1+exp (- x)) living^-1Output a is obtained, as follows.

Z=wx+b₁, a=f (z) (6)

Wherein, w ∈ R^S×D(D represents input data dimension and S represents hidden layer neuron number) and b₁∈R^S×1Generation respectively Weight and biasing of the table input layer to hidden layer.Further, initial input is similar in order to verify whether the reconstruct y for obtaining, right The a of acquisition enters row decoding, as follows：

T=va+b₂, y=f (t) (7)

Wherein, v ∈ R^D×S、b₂∈R^D×1Hidden layer is represented respectively to weight and the biasing of output layer.Initial input x and output Minimal reconstruction error between yCalculated.Especially, in order to find initial input some are special Fixed structure and dependency, sparse constraint and weight penalty term are added in network, and object function J is as follows：

Wherein, M is number of samples, x_mAnd y_mM-th input and output are represented respectively.λ representation severely punishes penalty parameter and η generations The weight of the sparse penalty term of table.Between j-th input neuron of i-th hidden neuron and l-1 layers that represent l layers Connection weight；Additionally,(target of hidden unit is averagely activated to represent r Function) andThe relative entropy of (the average activation primitive of hidden neuron i).

Such as formula (8), generally decline weight and the biasing that hidden layer is updated with back-propagation algorithm using gradient, so as to reality Existing optimization object function J.However, during carrying out objective function optimization again, model is more difficult to choose appropriate initial weight and partially Put, optimization process is easy to be absorbed in local extremum.In order to address an above-mentioned difficult problem, the present invention proposes to utilize improved artificial fish school intelligent Optimized algorithm substitutes gradient and declines and back-propagation algorithm, dilute using each layer of improved artificial fish school intelligent optimizing algorithm optimization The object function of thin autocoder.The reason for improvement is to represent the tracking performance of robust, fast receipts due to artificial fish-swarm algorithm Hold back speed and high global optimization ability.Specific embodiment is as follows：

If shoal of fish sum is n, each fish represents the difference and additivity value of a neutral net and any two fishes still A neutral net is represented, the parameters optimization of sparse autocoder includes two weight matrix [w] and [v] and threshold vector b； In current position, the food concentration of individual fish is Y=1/E, two individual fish X_pAnd X_qThe distance between as follows：

Wherein, output layer neuron number is N, w_ij(p) and w_ijQ () represents Artificial Fish X respectively_pAnd X_qParameter matrix I-th row, the element of jth row, v in [w]_ki(p) and v_kiQ () represents Artificial Fish X respectively_pAnd X_qParameter matrix [v] in row k, The element of the i-th row, b_i(p) and b_iQ () represents Artificial Fish X respectively_pAnd X_qThreshold vector b in the i-th row element, b_k(p) and b_k Q () represents Artificial Fish X respectively_pAnd X_qThreshold vector b in row k element, d_pqRepresent two individual fish X_pAnd X_qBetween away from From.

The present invention is studied and improves four biological behaviours of the shoal of fish, including foraging behavior, behavior of bunching, and following behavior And random behavior.

Foraging behavior：W (p) is the current state of individual fish, is then being randomly chosen a new shape within sweep of the eye State w (q), if food concentration Y_q＞ Y_p, w (p) is updated according to formula (10)；Otherwise, a new state is selected again Judge whether to meet advance condition, if being still unsatisfactory for advance condition after maximum attempts are executed, individual fish will be according to Formula (11) executes random behavior.

W (pnext)=w (p)+rand () × step+ χ × (w_ij(best)-w_ij(l)),if(Y_q≤Y_p) (11)

Wherein, w (pnext) is the next step state of individual fish, and rand () is equably to produce one between [0,1] at random Number.The present invention is proposed χ × (w_ij(best)-w_ij(l)) it is added in the more new formula of fish school location, it is called global optimum It is oriented to shoal of fish search terms.Wherein, w_ijAnd w (best)_ijL () represents current optimum Artificial Fish X respectively_bestWith random fish X_lParameter I-th row in matrix [w], the element of jth row.χ is golden ratio coefficient lambda and skewProduct, as follows：

Wherein,Represent the equally distributed random value of standard and interval isHereKnowable to formula (12), the new explanation of generation does not stay in identical position, but shifts an angle, so that Individual fish obtains more positional informationes in search space.

On the other hand, for slow, the present invention that overcomes the shortcomings of that Traditional Man fish-swarm algorithm has a later stage iterative convergence speed Propose to dynamically adjust visual field distance (V) and moving step length (M) using fuzzy logic system.In this designed system, The dispersion and iterationses in fish school location is used as input variable.V and M is used as output variable simultaneously.Dispersion measure expression is public Formula is as follows：

Wherein, shoal of fish sum represents current iteration for n and k.WithArtificial Fish X during kth time iteration is represented respectively_p With optimum Artificial Fish X_best.D represents dispersibility of the input variable of fuzzy system i.e. between the shoal of fish.Generally, the shoal of fish closer to When dispersibility lower, vice versa.In above-mentioned formula (13), Euclidean distance is used to measure between individual fish and optimum fish Dispersion.Additionally, calculating for convenience, iteration and dispersibility tolerance are normalized between 0 to 1, as follows：

K_Norm=k/K_max(14)

In formula, K_NormAnd D_NormCurrent iteration number of times k and dispersibility D is carried out the fuzzy system after standard normal respectively System |input paramete；K_maxFor maximum iteration time；D_minAnd D_maxThe minima of dispersibility and maximum in respectively each iteration.

It should be noted that above-mentioned updating location information rule is equally applicable to bunch behavior and following behavior.In order to Each fish, searches for optimal solution by four kinds of behavioral pattern of execution foregoing description.Then, best behavioral pattern is selected Update current state.Finally need to be updated bulletin board, i.e. the food concentration of the optimum state of Artificial Fish and maximum is remembered On bulletin board, each fish is updated after movement for record, then compares bulletin board and the new state for producing.If artificial The current state of fish is better than bulletin board, and the value on bulletin board will be substituted.

After each sparse autocoder has been trained, reconstruction of layer is removed and the feature extracted be stored in hiding Layer, is then used as the input of next hidden layer, so as to produce higher order feature.Further, multiple non-by stacking Supervise feature learning layer to build the sparse autocoder of storehouse.Finally, random forest grader be concatenate to storehouse sparse from Last layer of dynamic encoder, when train grader when to study to high-order feature carry out the fine setting for having supervision, then pass through Maximum voting rule differentiates the class label of test set.

Fig. 1 is the united classification hyperspectral imagery flow chart of the empty spectrum of a width, in the trilateral filter for proposing, by synthesis The spectrum and spacial proximity of consideration high spectrum image is so as to producing spectral-spatial information characteristics collection.The smoothed image of acquisition can To effective filter out Gaussian noise, speckle noise and impulsive noise etc., while retaining the detailed information of image.Additionally, of the invention Non-supervisory training is carried out to the smoothed image for obtaining using the sparse autocoder of improved storehouse and random forest grader With the fine setting and classification for having supervision.The deep learning machine model for being proposed can extract the abstract and useful of data layer by layer High-order character representation, so that improve the classification performance of high-spectral data.

Fig. 2 is the flow chart of the sparse autocoder of a width artificial fish school algorithm.In improved deep learning model In, global optimum is oriented to shoal of fish search terms and fuzzy logic ordination is incorporated in fish school behavior pattern, realize optimization aim letter Number J.The proposition of the method can keep the multiformity of the shoal of fish and while avoid algorithm in the slow deficiency of later stage iteration convergence, so as to It is effectively prevented from deep learning model to be easy to be absorbed in local extremum during optimization object function, so as to obtain global optimum Solution.

Fig. 3 is an Artificial Fish model,WithFor two any fishes, X_pFor current manual fish, step is mobile step Long, visual is its sensing range, is then being randomly chosen a new Artificial Fish X within sweep of the eye_qIf, food concentration Value Y_q＞ Y_p, Artificial Fish X_pCurrent state w (p) be updated；If Artificial Fish X_qNew state w (q) unlike current state more Excellent, then continue the other positions that makes an inspection tour in sensing range, judge whether to meet advance condition.If executing maximum trial Still advance condition is unsatisfactory for after number of times, and arbitrarily shifting is moved a step by individual fish.Generally, the final goal of shoal of fish movement is to pass through Optimum behavioral pattern is executed so as to obtaining optimal solution.

Ginseng Fig. 4 (a)-Fig. 4 (c), the figure are used to input and the output variable for showing fuzzy logic system.In system, fish The dispersion and iterationses of group position is used as input variable.Field range and moving step length are used as output variable simultaneously.Fig. 4 L in (a)-Fig. 4 (c), M, H, ML and MH represent respectively low, in, high, in low and middle high level.Generally during earlier iterations, depending on Wild scope and moving step length should take higher value so that searching for wider and guaranteeing global optimization.Then, with iteration Carry out, field range and moving step length should gradually reduce to guarantee that algorithm executes optimum in the range of neighbouring globally optimal solution Change, so as to obtain high precision solution.

Fig. 5 is that the sky based on trilateral filter and the sparse autocoder of storehouse composes united classification hyperspectral imagery frame Frame.

First, using propose trilateral filter improve picture quality, will spatial information be incorporated in spectral domain while Reduce the degradation phenomena caused by noise.Further, smooth image is extracted using the sparse autocoder of improved storehouse High-order feature.Finally, by the abstract high-order characteristic action for obtaining in random forest grader, obtained by main ballot criterion Take final classification results.

Fig. 6 (a)-Fig. 6 (h) is the real hyperspectral image data of a width, and this group of data are by AVIRIS sensors in 1992 Year is obtained in the farm overhead of the Indiana, USA northwestward.AVIRIS packets contain 224 wave bands, its spatial resolution Be 20m and spectral coverage be 0.2～2.4 μm.20 water absorption bandses and 4 noise wave bands are being removed, remaining 200 ripples Section is used for classification experiments.View picture figure size is 144x144 pixel, comprising the 16 classes ground not waited from 20 to 2468 pixels Thing.In order that experimental analysiss are more significantly, seven classes are removed and remaining nine classes large sample atural object quilt with regard to the atural object of small sample Retain.Fig. 6 (a) represents that the true atural object distribution of scene, Fig. 6 (b)-(d) represent initial spectrum data set, come from bilateral filtering The empty spectrum union feature of the empty spectrum union feature of device and the trilateral filter for coming from proposition is respectively fed to support vector machine The classification chart (being expressed as OS-SVM, BF-SVM, and TF-SVM) of grader.Fig. 6 (e)-(g) represents initial spectrum data Collection, the empty spectrum union feature for coming from two-sided filter and the empty spectrum union feature for coming from trilateral filter are respectively fed to The sparse autocoder of storehouse and random forest grader (being expressed as OS-SSARF, BF-SSARF, and TF-SSARF). Finally, Fig. 6 (h) is represented and is come from the empty spectrum union feature of trilateral filter and be fed to the sparse autocoder of improved storehouse With random forest grader (as TF-ISSARF).From fig. 6, it can be seen that compared to contrast algorithm, the TF- for being proposed ISSARF methods obtain best classification chart (i.e. Fig. 6 (h)).

Fig. 7 (a)-Fig. 7 (h) is the real hyperspectral image data of a width, and this group of data are equally by AVIRIS sensors Acquire, it shoots Kennedy Space Center of the U.S. overhead in 1996.Data constitute feature and Indian farm number Consistent according to collection.Through data prediction, retaining 176 wave bands carries out classification experiments.Image size is 614 × 512 pixels, and space is divided Resolution 18m.In Fig. 7 (a), the true atural object distribution of scene is we illustrated.Fig. 7 (b)-(h) is respectively and contrasts algorithm and institute The classification chart of the algorithm of proposition.Consistent with expected resultss, the TF-ISSARF methods for being proposed obtain best classification chart (i.e. Fig. 7 (h)) and it be distributed substantially close to real atural object.

With reference to Fig. 8 (a)-Fig. 8 (b), research and analysis spatial parameter of the general classification accuracy with trilateral filter, Magnitude parameters and the variability of pulse parameter value, in Fig. 8 (a), for Indiana farm data, work as δ_s=7, δ_r= 0.4, and δ_dThe TF-ISSARF sorting techniques proposed when=8 obtain optimum general classification accuracy.In for Kennedy Space Calculation evidence, understands to work as δ from Fig. 8 (b)_s=8, δ_r=0.4, and δ_dWhen=7, optimum general classification accuracy is obtained.

With reference to Fig. 9 (a)-Fig. 9 (b), the neighborhood window size of research and analysis trilateral filter is accurate to general classification The impact of degree.In order to obtain the window size of optimum, for Indiana farm data, fixed δ_s=7, δ_r=0.4, and δ_d= 8, analysis general classification accuracy is with the change of window size.Knowable to Fig. 9 (a), as window size w=7, farm data Obtain best classification results.For Fig. 9 (b), as window size w=9, Kennedy Space Center's data obtain best dividing Class result.

Above-mentioned for the present invention especially exemplified by embodiment, be not limited to the present invention.What the present invention was provided is filtered based on three sides The hyperspectral image classification method of device and the sparse autocoder of storehouse is equally applicable to other non-high spectrum images of classifying.? Without departing from the spirit and scope of the invention, a little adjustment and optimization can be done, with protection scope of the present invention with claim It is defined.

Claims (6)

1. a kind of based on trilateral filter and the image classification method of the sparse autocoder of storehouse, it is characterized in that：

First, smooth image is obtained using trilateral filter, while the spectral-spatial feature of the pixel for extracting described image Filter Gauss, speckle and the impulsive noise of degraded image；

Secondly, high-order feature extraction is carried out using improved storehouse sparse autocoder；

Finally, carried out supervising trim network and classification using random forest grader.

2. according to claim 1 based on trilateral filter and the image classification method of the sparse autocoder of storehouse, its It is characterized in that：The pixel of the trilateral filter more new formula is：

$I^\left(t\right)=\frac{1}{W_t}\underset{v\∈S_p}{\Σ}\exp (-\frac{\left|\right|t-v|{|}^2}{2{\mathrm{\δ}}_s^2}-\frac{\left|\right|J\left(t\right)-J\left(v\right)|{|}^2}{2{\mathrm{\δ}}_r^2}-\frac{ROAD\left|\right|I\left(t\right)-I\left(v\right)|{|}^2}{2{\mathrm{\δ}}_d^2})I\left(v\right)$

Wherein, t and v represent the coordinate of pixel, S respectively_pRepresent the set of the neighborhood of pixel t, δ_s、δ_rAnd δ_dSpace is respectively controlled Parameter, magnitude parameters and pulse parameter, I (t) is the center pixel of initial pictures, and I (v) is the pixel in neighborhood window, J (t) It is the pixel in center pixel and neighborhood window based on reference picture respectively with J (j), W_tIt is spacial proximity weightingSimilarity is weightedWeight with pulseProduct, W_tIt is further represented as：

$W_t=\underset{v\∈S_p}{\mathrm{\Σ}}\exp (-\frac{\left|\right|t-v|{|}^2}{2{\mathrm{\δ}}_s^2}-\frac{\left|\right|J\left(t\right)-J\left(v\right)|{|}^2}{2{\mathrm{\δ}}_r^2}-\frac{ROAD\left|\right|I\left(t\right)-I\left(v\right)|{|}^2}{2{\mathrm{\δ}}_d^2})$

3. according to claim 2 based on trilateral filter and the image classification method of the sparse autocoder of storehouse, its It is characterized in that：When executing pulse and weighting, determine that using grade sequence absolute difference sample point is the marginal point of image or by arteries and veins The pixel of sound pollution is rushed, corresponding computing formula is as follows：

D [t, v]=| I (t)-I (v) |

r_f(t, v)=f^thsmallest d[t,v]

${\mathrm{ROAD}}_g\left(t,v\right)=\underset{f=1}{\overset{g}{\Σ}}{r}_f(t,v)$

Wherein, r_f(t, v) represent increasing arrangement antipode value d [t, v], i.e., from small to large add up choose f d [t, v] value, 2 ≤g≤w²- 2, w are window sizes, in equationIn, ROAD statistics is used to measure center pixel Proximity with g-th neighborhood territory pixel.

4. according to claim 1,2 or 3 based on trilateral filter and the image classification side of the sparse autocoder of storehouse Method, is characterized in that described carrying out high-order feature extraction using the sparse autocoder of improved storehouse and specifically including：

Sparse autocoder is attempted extracting the input data for hiding feature so that the reconstructed vector of decoding layer is similar to input layer, That is input vector x ∈ R^DIt is mapped to hidden layer and produces a linear composition z ∈ R^S, then adopt a nonlinear activation letter Number f (x)=(1+exp (- x))^-1Output a is obtained, as follows

Z=wx+b₁, a=f (z)

Wherein, w ∈ R^S×DAnd b₁∈R^S×1Input layer is represented respectively to weight and the biasing of hidden layer, D represent input data dimension, S represents hidden layer neuron number, in order to verify whether that the reconstruct y for obtaining is similar to initial input, a for obtaining is translated Code, as follows：

T=va+b₂, y=f (t)

Wherein, v ∈ R^D×S、b₂∈R^D×1Hidden layer is represented respectively to weight and the biasing of output layer, initial input x and reconstruct y it Between minimum errorCalculated；In order to find some ad hoc structures and the correlation of initial input Property, sparse constraint and weight penalty term are added in network, and object function J is as follows：

$J=\frac{1}{2}\underset{m=1}{\overset{M}{\Σ}}\left|\right|y_m-x_m|{|}_2^2+\frac{\mathrm{\λ}}{2}\underset{l}{\mathrm{\Σ}}\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{\left(w_{i,j}^{\left(l\right)}\right)}^2+\mathrm{\η}\underset{i=1}{\overset{S}{\Σ}}KL\left(r\right|\left|{\stackrel{\‾}{r}}_i\right)$

Wherein, M is number of samples, x_mAnd y_mM-th input and output are represented respectively, and λ representation severely punishes penalty parameter, and η represents sparse The weight of penalty term,Connection weight between j-th input neuron of i-th hidden neuron and l-1 layers that represent l layers Weight；Additionally,Represent the average activation primitive r of target of hidden unit and hidden Hide the average activation primitive of neuron i

Using the object function of the sparse autocoder of each layer of improved artificial fish school intelligent optimizing algorithm optimization, concrete scheme As follows：

If shoal of fish sum is n, each fish represents a neutral net and the difference and additivity value of any two fishes are still represented One neutral net, the parameters optimization of sparse autocoder include two weight matrix [w] and [v] and threshold vector b；Working as Front position, the food concentration of individual fish is Y=1/E, two individual fish X_pAnd X_qThe distance between as follows：

$d_{pq}=\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\underset{j=1}{\overset{D}{\mathrm{\Σ}}}{\[w_{ij}\left(p\right)-w_{ij}\left(q\right)\]}^2+\underset{k=1}{\overset{N}{\mathrm{\Σ}}}\underset{i=1}{\overset{S}{\mathrm{\Σ}}}{\[v_{ki}\left(p\right)-v_{ki}\left(q\right)\]}^2+\underset{i=1}{\overset{S}{\mathrm{\Σ}}}{\[b_i\left(p\right)-b_i\left(q\right)\]}^2+\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{\[b_k\left(p\right)-b_k\left(q\right)\]}^2$

Wherein, output layer neuron number is N, w_ij(p) and w_ijQ () represents Artificial Fish X respectively_pAnd X_qParameter matrix [w] in i-th Row, the element of jth row, v_ki(p) and v_kiQ () represents Artificial Fish X respectively_pAnd X_qParameter matrix [v] in row k, i-th row unit Element, b_i(p) and b_iQ () represents Artificial Fish X respectively_pAnd X_qThreshold vector b in the i-th row element, b_k(p) and b_k(q) generation respectively Table Artificial Fish X_pAnd X_qThreshold vector b in row k element, d_pqRepresent two individual fish X_pAnd X_qThe distance between.

Foraging behavior：W (p) is the current state of individual fish, then being randomly chosen new state w (q) within sweep of the eye, if Food concentration Y_q＞ Y_p, w (p) is according to formula It is updated；Otherwise, select a new state again to judge whether to meet advance condition, if executing maximum trial time Still advance condition is unsatisfactory for after number, and individual fish will be according to formula w (pnext)=w (p)+rand () × step+ χ × (w_ij (best)-w_ij(l)),if(Y_q≤Y_p) random behavior is executed,

Wherein, w (pnext) is the next step state of individual fish, and rand () is equably to produce a random number between [0,1], w_ijAnd w (best)_ijL () represents current optimum Artificial Fish X respectively_bestWith random fish X_lParameter matrix [w] in the i-th row, jth row Element, χ is golden ratio coefficient lambda and skewProduct representation be：

Wherein,Represent the equally distributed random value of standard and interval is

The visual field is dynamically adjusted using fuzzy logic system apart from V and moving step length M, the dispersion in fish school location and iteration time , used as input variable, while V and M is used as output variable, dispersion measure expression formula is as follows for number：

$D=\frac{1}{n}\underset{p=1}{\overset{n}{\Σ}}\sqrt{{(X_p^k-X_{best}^k)}^2}$

Wherein, shoal of fish sum is n, and k represents current iteration,WithArtificial Fish X during kth time iteration is represented respectively_pMost Excellent Artificial Fish X_best, D represents dispersibility of the input variable of fuzzy system i.e. between the shoal of fish, and iteration and dispersibility are measured by normalizing Change between 0 to 1, as follows：

K_Norm=k/K_max

$D_{Norm}=\left\{\begin{array}{cc}\frac{D-D_{min}}{D_{\max }-D_{\min }},& D_{\min }\≠D_{max}\\ 0,& D_{\min }=D_{max}\end{array}\right.$

In formula, K_NormAnd D_NormCurrent iteration number of times k and dispersibility D is respectively carried out the fuzzy system after standard normal defeated Enter parameter；K_maxFor maximum iteration time；D_minAnd D_maxThe minima of dispersibility and maximum in respectively each iteration.

5. according to claim 1,2 or 3 based on trilateral filter and the image classification side of the sparse autocoder of storehouse Method, is characterized in that：Described image is high spectrum image.

6. according to claim 4 based on trilateral filter and the image classification method of the sparse autocoder of storehouse, its It is characterized in that：Described image is high spectrum image.