CN106599848A - Depth visual feature and support vector machine-based terrain texture recognition algorithm - Google Patents

Abstract

The present invention provides a depth visual feature and support vector machine-based terrain texture recognition algorithm. The algorithm includes the following steps that: step 1, the pictures of different kinds of unstructured roads are acquired, a part of each kind of pictures are stored into a training set, and the other part of each kind of pictures are stored into a test set, and the training set and the test set constitute a terrain texture database; step 2, the AlexNet convolutional neural network model is adopted to obtain the depth visual features of all the images of the terrain texture database, and the depth visual features of all the images of the terrain texture database are stored; step 3, tags are defined for the different unstructured roads in the established terrain texture database, and a support vector machine algorithm is adopted to analyze the depth visual features of the training set to obtain a decision function and analyze the depth visual features of the test set to obtain the predictive tags and classification accuracy of the test set; and step 4, whether the accuracy is higher than a set value is judged, if the accuracy is higher than the set value, the terrain texture database is saved, otherwise, the pictures of different kinds of unstructured roads are re-acquired.

Description

Based on deep vision feature and the satellite imagery recognizer of SVMs

Technical field

It is particularly a kind of based on deep vision feature and the ground of SVMs the present invention relates to a kind of mode identification technology Table texture recognizer.

Background technology

The intelligent vehicles technology is an emerging subject, has merged computer, modern sensing, information fusion, communication, manually Intelligence and the technology such as automatically control, be widely used in military affairs, civil area.In civil field, it offers convenience for driver While, the generation of emergency situations can be effectively prevented from, reduce the generation of traffic accident.The ground floor technology of the technology is intelligence Energy perception/early warning system, is obtained to vehicle itself, the surrounding environment of vehicle traveling and driver using various sensor informations Platform itself is perceived, and early warning information is sent if necessary, is DAS (Driver Assistant System) and Vehicular automatic driving system this rear two-layer skill The basis of art.

During due to travelling in intelligent vehicle country not only out of doors, it is necessary to first earth's surface is identified, road is analyzed Change, then carry out different operations so as to realize the independent navigation of vehicle.Road substantially can be divided into structured road with Unstructured road, for the difference of road structure, the research of Surface classification is also different.In terms of feature extraction, structure Color, the Texture eigenvalue for changing Road Surface is relatively stable, unified, is convenient for targetedly studying, main by recognizing road The shape of route and road sign carrys out modeling analysis, and current studies comparative maturity and have larger progress.And non-structural The Surface classification for changing road is larger due to being affected by the external environment, such as under sunlight, sleet, season equity outdoor environment, road table The result of classification is just corresponding poor, is still within conceptual phase and extremely challenging, needs constantly developing and perfect badly.

The content of the invention

It is an object of the invention to provide a kind of satellite imagery identification of base based on deep vision feature and SVMs Algorithm.The present invention can preferably eliminate sunlight, sleet, season in the satellite imagery database that known open air collects Impact of the reciprocity outdoor environment to unstructured road, extracts the deep vision feature with robustness and ability to see things in their true light, and enters Row classification of road, classification accuracy is high.

It is a kind of based on deep vision feature and the satellite imagery recognizer of SVMs, comprise the following steps：

Step 1, obtains the picture of unstructured road not of the same race, and it is another that a part for every class picture is stored in into training set Part is stored in test set, and training set and test set constitute satellite imagery database；

Step 2, using AlexNet convolutional neural networks models, the depth for obtaining all images of satellite imagery database is regarded Feel feature and preserve；

Step 3, in the satellite imagery database set up, to unstructured road not of the same race label is defined respectively, is adopted The deep vision feature of algorithm of support vector machine analyzing and training collection obtains decision function, and the deep vision feature for analyzing test set is obtained To the prediction label and classification accuracy of test set；

Whether step 4, judging nicety rate is higher than setting value, and if higher than setting value satellite imagery database is retained；Otherwise, From the new picture for obtaining unstructured road not of the same race.

The present invention compared with prior art, with advantages below：(1) the invention has the advantages that being based on " Caffe " depth Framework is practised, the depth for obtaining all input pictures of satellite imagery database extraction using AlexNet convolutional neural networks models is regarded Feel feature, single input picture exports one 4096 characteristic vector tieed up through 5 convolutional layers and 2 full articulamentums, preferably Sunlight, sleet, season equity impact of the outdoor environment to unstructured road are eliminated, is extracted with robustness and ability to see things in their true light Deep vision feature；(2) using SVMs (Support Vector Machine, SVM) algorithm to deep vision feature It is analyzed, and then reaches the purpose of identification satellite imagery image, with very high recognition accuracy.

The present invention is described further with reference to Figure of description.

Description of the drawings

Fig. 1 is the embodiment of the present invention based on deep vision feature and the flow process of the satellite imagery recognizer of SVMs Schematic diagram.

Fig. 2 is the cross-country satellite imagery database sample schematic diagram of Institutes Of Technology Of Nanjing.

Fig. 3 is convolutional neural networks model structure schematic diagram.

Specific embodiment

It is a kind of based on deep vision feature and the satellite imagery recognizer of SVMs, it is characterised in that include with Lower step：

Step 1, obtains the picture of unstructured road not of the same race, and it is another that a part for every class picture is stored in into training set Part is stored in test set, and training set and test set constitute satellite imagery database；

Step 2, using AlexNet convolutional neural networks models, the depth for obtaining all images of satellite imagery database is regarded Feel feature and preserve；

Step 3, in the satellite imagery database set up, to unstructured road not of the same race label is defined respectively, is adopted The deep vision feature of algorithm of support vector machine analyzing and training collection obtains decision function, and the deep vision feature for analyzing test set is obtained To the prediction label and classification accuracy of test set；

Whether step 4, judging nicety rate is higher than setting value, and if higher than setting value satellite imagery database is retained；Otherwise, From the new picture for obtaining unstructured road not of the same race.

In step 2, each image exports deep vision characteristic vector through 5 convolutional layers and 2 full articulamentums.

Algorithm of support vector machine described in step 4 is specially：

Step 4.1.1, the hyperplane set up in the data space of n dimensions, its expression formula is

ω^TX+b=0

ω is n-dimensional vector, and b is constant；

Step 4.1.2, structural classification function causes hyperplane effectively to separate two class data, and classification function is：

F (x)=ω^Tx+b

Step 4.1.3, determines the target of largest interval grader, and expression formula is：

Step 4.1.4, by maximization problems minimization problem is converted into, and targeted transformation is：

Step 4.1.5, using Lagrange dualities Lagrange dual variables α are introduced, and are added to each constraints A upper Lagrange multiplier, constraints is dissolved in object function, and problem is attributed to：

s.t.0≤α_i≤ C, i=1,2 ..., n

Step 4.1.6, in the case where two inequality are as constraints extreme value, existence anduniquess optimal solution are solvedThen equation Solution be：

Step 4.1.7, the optimal hyperlane equation for finally giving is：

Deep vision feature in the step 4 using algorithm of support vector machine analyzing and training collection obtains decision function, has Body process is：

Step 4.2.1, in the satellite imagery database set up, to the different labels of unstructured road distribution not of the same race；

Step 4.2.2, the deep vision using the svmtrain function pair training sets in libsvm-weights-3.20 is special Levy and be trained, obtain the model for training；

Step 4.2.3, with the deep vision of the svmpredict Functional Analysis test sets in libsvm-weights-3.20 Feature obtains the prediction label and classification accuracy of test set.

Embodiment

With reference to Fig. 1, the implementation steps of the present embodiment are as follows：

Step 1, in the cross-country satellite imagery database of Institutes Of Technology Of Nanjing, extracts 600 images, wherein three class non-structural Change the picture of road, the respectively width of soil (dirt) road 200, the width of sandstone (gravel) road 200, the width of water (water) 200, it is each The each 100 training set pictures (traindata) of class picture and 100 test set pictures (testdata) are respectively used to train and survey Examination, sets up satellite imagery database；

Step 2, using AlexNet convolutional neural networks models, using satellite imagery database all 600 inputs is extracted The deep vision feature of image, saves as text document, the characteristic vector of totally 600 4096 dimensions.Each input picture is through 5 volumes Lamination and 2 full articulamentums export one 4096 characteristic vector tieed up；

Step 3, using matlab the data in above-mentioned text document are read in, and data are recombinated, and obtain 600*4096 The matrix alldata of size is input into as algorithm of support vector machine；

Step 4, it is two matrixes of training set traindata and test set testdata that matrix alldata is divided to, wherein instructing Practice concentration to include before soil (dirt) road in deep vision characteristic, sandstone (gravel) road, each class of water (water) 100, soil (dirt) road in deep vision characteristic, sandstone (gravel) road, water (water) are included in test set often Latter 100 of one class；

Step 5, in the satellite imagery database set up, to the label label of soil, sandstone, water mark is respectively defined as Sign 1, label 2, label 3；

Step 6, svmtrain functions are used to be trained the data of training set traindata, obtain the mould for training Type model；

Step 7, svmpredict functions are used to analyze the deep vision feature of test set and obtain the prediction label of test set With predict_label and classification accuracy accuracy；

Step 8, above-mentioned classification accuracy is exported as a result；

In step 1, it is necessary first to set up satellite imagery database, the colored RGB that 600 pixels are 25*25*3 is obtained Image set.

As shown in Fig. 2 the earth's surface of three class unstructured roads is larger due to being affected by the external environment, for example sunlight, sleet, Season equity outdoor environment factor, the result of road table sort is just corresponding poor, so needing by convolutional neural networks training ground Table texture database obtaining sane, strong deep vision feature, so as to preferably carry out Surface classification.

Convolutional neural networks compare that traditional neural network is more complicated, and neuron shares connection weight, per two-layer neuron with Locally-attached mode is attached, and the time and spatially using dimensionality reduction sample, take full advantage of data.

For image recognition, using the image of equal resolution as input, the output conduct of each afterwards layer of neutral net Next layer of input, the neuron in network is only connected with adjacent neurons, so as to form local sensing.Neuron can be at this The local feature of image is extracted in the structure of sample, and the local feature for extracting can also characterize general image.

Satellite imagery database is obtained using AlexNet convolutional neural networks models and extracts all 600 input pictures Deep vision feature, single input picture exports one 4096 characteristic vector tieed up through 5 convolutional layers and 2 full articulamentums. With reference to Fig. 3, step 2 is concretely comprised the following steps：

Step 2.1, ground floor convolution：The convolution kernel of 5 × 5 sizes 96, each GPU is upper 48.Ground floor maximum pond： 2 × 2 core；

Step 2.2, second layer convolution：3 × 3 convolution kernels 256, each GPU is upper 128.Second layer maximum pond：2×2 Core；

Step 2.3, third layer convolution：It is to be connected entirely with last layer, the convolution kernel of 3*3 384.Assign to two GPU last 192；

Step 2.4, the 4th layer of convolution：3 × 3 convolution kernel 384, two GPU are each 192.The layer is connected with last layer Not through pond layer；

Step 2.5, layer 5 convolution：3 × 3 convolution kernel 256, two GPU were last 128.Layer 5 maximum pond： 2 × 2 core；

Step 2.6, ground floor connects entirely：4096 dimension, by layer 5 maximum pond output be connected to become one it is one-dimensional to Amount, as the input of this layer；

Step 2.7, the second layer connects entirely：4096 dimensions, use as output for follow-up SVM classifier.

The information of image both remained by convolutional layer in step 2 or the scale of parameter had been reduced.Convolution is a little square Battle array is acted on picture element matrix, defines the convolution kernel of a m*n, and convolution kernel acts on each position of image according to fixed step size Put, the neuron number for so connecting pixel reduces, and simplifies training process.

The complexity for both having reduced calculating by pond layer (maximum pond layer) in step 2 in turn ensure that the correct of result. Using a filter and the stride of an equal length, maximum is taken to each rectangle in filter and is recorded to another square Battle array.

By pond layer each neuron can be connected with next layer each neuron in step 2.Through convolution The Network Capture feature of the consistency of high-order, these profile feedbacks to full articulamentum, by full articulamentum pair after layer and pond layer These features are classified.Process through full connection hidden layer and the conversion of output layer information and calculating, just complete and once learn Forward-propagating processing procedure, final result outwardly exports by output layer.

ReLU functions are employed in step 2, its expression formula is：

F (x)=max (0, x)

ReLU is a kind of nonlinear function, and form is, if calculated value is less than 0, to make it be equal to 0, and its residual value is constant., ReLU nonlinear functions are simply compared, it is not necessary to which complicated data are calculated, so that the neutral net of training can be faster Ground convergence.

Employ feature normalization in step 2, i.e., in order that the adjacent position of same characteristic pattern and different characteristic figure The feature of same position evenly, needs the feature to same position to carry out division normalization operation.If a_x,yFor former neuron Activation, b_x,yFor the activation of new neuron, the calculating of the newly-generated activation of neuron its expression formula is：

Adopt Dropout methods in step 2 to prevent network from over-fitting occur.Dropout is referred to will with 50% probability Each hidden layer neuron zero setting, prevents interdepending between feature, can be effectively reduced test errors.

Step 2, using AlexNet convolutional neural networks models, using satellite imagery database all 600 inputs is extracted The deep vision feature of image, saves as text document, the characteristic vector of totally 600 4096 dimensions.Each input picture is through 5 volumes Lamination and 2 full articulamentums export one 4096 characteristic vector tieed up；

In step 3, call importdata functions to read in the data in above-mentioned text document using matlab, call Reshape function pair data are recombinated, and the matrix alldata for obtaining 600*4096 sizes is defeated as algorithm of support vector machine Enter；

In step 6, according to feature traindata of category trainlabel and training sample for being training sample, call The svmtrain functions of libsvm-weights-3.20 are used to be trained the data of training set traindata, are trained Good model model, expression is：

Model=svmtrain ([], trainlabel, traindata, '-t 0').

Wherein trainlabel is the category of training sample, and traindata is the feature of training sample, and-t is kernel function class Type, 0 represents linear kernel function, 1 representative polynomial kernel function, 2 expression RBF (radial direction base) kernel functions, 3 expression sigmoid core letters Number, takes t=1.

In step 7, what SVMs considered is two classification problems, and data point is represented with x, and this is a n dimension Vector, and classification is represented with y, can take 1 or -1, and two different classes are represented respectively.The target of study is that construction one is sentenced Other function, test data is classified as correctly as possible. and a linear classifier seeks to find one in the data space of n dimensions Individual hyperplane, its equation can be expressed as：

ω^TX+b=0

。

Defining classification function is：

F (x)=ω^Tx+b。

The hyperplane that causes of classification function effectively separates two class data, i.e. margin maximization between two classes, maximum The target of Margin Classification device is：

Again maximization problems is converted into into minimization problem, targeted transformation is

Maximum class interval is asked so as to the problem for seeking classification function is transformed into, the optimum to ω and b is then converted to then Change problem, i.e. convex quadratic programming problem.

Lagrange dual variables α are introduced using Lagrange dualities, to each constraints one is added Lagrange multipliers, constraints is dissolved in object function, and problem is attributed to

s.t.0≤α_i≤ C, i=1,2 ..., n

Extreme value, existence anduniquess optimal solution are solved in the case where two inequality are as constraintsThen non trivial solution is：

Step 4.1.7, the optimal hyperlane equation for finally giving is：

In step 7, according to the label testlabel and the eigenmatrix testdata of test sample of test sample, call The svmpredict functions of libsvm-weights-3.20 are used to analyze the deep vision feature of test set and obtain the pre- of test set Mark label and predict_label and classification accuracy accuracy, expression is：

[predict_label, accuracy, prob_estimates]=svmpredict (testlabel, testdata,model)。

Wherein testlabel is the label of test sample, and testdata is the eigenmatrix of test sample.As a result export Predict_label is the label that prediction is obtained, and accuracy is classification accuracy.

In step 8, above-mentioned classification accuracy is exported as a result, it is non-to three classes that table 1 illustrates the inventive method The classification accuracy of structured road, the accuracy rate of gravel road is up to 95%, and the accuracy rate of muddy and ruthed lane and water this two class then reaches 100% (completely errorless).Satellite imagery recognizer based on deep vision feature and SVMs can be on the ground set up Discrimination on table texture database reaches 98.33%.Additionally, also directly with SVMs to the database used by the present invention Classified, discrimination is only 62%, classification accuracy is improve 31.33% by algorithm proposed by the present invention by comparison.This The algorithm that invention is proposed has very high validity, accuracy and robustness, and this is for satellite imagery identification and unmanned skill Art has great significance.

Table 1

Unstructured road species	Muddy and ruthed lane	Gravel road	Water
				The correct picture number of prediction label result	100	95	100
Classification accuracy	100%	95%	100%

Claims (4)

1. it is a kind of based on deep vision feature and the satellite imagery recognizer of SVMs, it is characterised in that including following Step：

Step 1, obtains the picture of unstructured road not of the same race, and a part for every class picture is stored in into training set another part Test set is stored in, training set and test set constitute satellite imagery database；

Step 2, using AlexNet convolutional neural networks models, the deep vision for obtaining all images of satellite imagery database is special Levy and preserve；

Step 3, in the satellite imagery database set up, defines respectively label, using support to unstructured road not of the same race The deep vision feature of vector machine Algorithm Analysis training set obtains decision function, and the deep vision feature for analyzing test set is surveyed The prediction label and classification accuracy of examination collection；

Whether step 4, judging nicety rate is higher than setting value, and if higher than setting value satellite imagery database is retained；Otherwise, from new Obtain the picture of unstructured road not of the same race.

2. method according to claim 1, it is characterised in that in step 2, each image is through 5 convolutional layers and 2 Full articulamentum output deep vision characteristic vector.

3. method according to claim 1, it is characterised in that the algorithm of support vector machine described in step 4 is specially：

Step 4.1.1, the hyperplane set up in the data space of n dimensions, its expression formula is

ω^TX+b=0

ω is n-dimensional vector, and b is constant；

Step 4.1.2, structural classification function causes hyperplane effectively to separate two class data, and classification function is：

F (x)=ω^Tx+b

Step 4.1.3, determines the target of largest interval grader, and expression formula is：

$max\frac{1}{\left|\right|\mathrm{\ω}\left|\right|}s.t.y_i({\mathrm{\ω}}^T{x}_i+b)\≥1,i=1,2,\mathrm{...},n$

Step 4.1.4, by maximization problems minimization problem is converted into, and targeted transformation is：

$min\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^{2}s.t.y_i({\mathrm{\ω}}^T{x}_i+b)\≥1,i=1,2,\mathrm{...},n$

Step 4.1.5, using Lagrange dualities Lagrange dual variables α are introduced, and to each constraints one is added Individual Lagrange multipliers, constraints is dissolved in object function, and problem is attributed to：

$\max W\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{i,j=1}{\overset{n}{\&Sigma;}}{y}_i{y}_j{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j<x_i,x_j>$

s.t.0≤α_i≤ C, i=1,2 ..., n

$\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i{y}_i=0;$

Step 4.1.6, in the case where two inequality are as constraints extreme value, existence anduniquess optimal solution are solvedThen non trivial solution For：

${\mathrm{\&omega;}}^{*}=\underset{i=1}{\overset{n}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}{x}_i$

$b^{*}=y_j\underset{i=1}{\overset{n}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i<x_i,x_j>$

Step 4.1.7, the optimal hyperlane equation for finally giving is：

$f\left(x\right)=\mathrm{sgn}({\left({\mathrm{\&omega;}}^{*}\right)}^{T}x+b^{*})=\mathrm{sgn}(\underset{i=1}{\overset{n}{\&Sigma;}}{\mathrm{\&alpha;}}_i^{*}{y}_i{x}_i^{*}x+b^{*}).$

4. method according to claim 1, it is characterised in that using algorithm of support vector machine analysis instruction in the step 4 The deep vision feature for practicing collection obtains decision function, and detailed process is：

Step 4.2.1, in the satellite imagery database set up, to the different labels of unstructured road distribution not of the same race；

Step 4.2.2, is entered using the deep vision feature of the svmtrain function pair training sets in libsvm-weights-3.20 Row training, obtains the model for training；

Step 4.2.3, with the deep vision feature of the svmpredict Functional Analysis test sets in libsvm-weights-3.20 Obtain the prediction label and classification accuracy of test set.