CN113298143A - Foundation cloud robust classification method - Google Patents

Abstract

The invention relates to a foundation cloud robust classification method which comprises two parts, wherein the first part is used for feature extraction and converts an image into a feature vector y; the second part is used to classify the input feature vector y. The invention relates to a weighted sparse representation-based convolutional neural network feature fusion cloud classification method, which is characterized in that two convolutional neural networks are extracted to serve as a dictionary for weighted sparse representation so as to improve the operation efficiency; by weighting sparse representation classification, the robustness of the system can be improved under the condition of occlusion, and the performance better than that of a single convolutional neural network can be obtained by fusing two convolutional neural networks.

Description

Foundation cloud robust classification method

Technical Field

The invention belongs to the technical field of classification of foundation cloud pictures, and particularly relates to a foundation cloud robust classification method.

Background

Cloud is an important weather phenomenon, and a reliable cloud observation technology has important significance for work such as climate research, weather analysis and weather forecast. The foundation cloud observation is an important cloud observation mode, can reflect the microstructure information of the cloud, makes up the defects of satellite observation, and can provide more comprehensive data for the cloud observation related application by fully applying the foundation cloud observation information. In the foundation cloud observation, the foundation cloud image classification technology is the key for realizing the foundation cloud observation, and the application of the technology can not only liberate observers from heavy cloud observation work, but also improve the accuracy and timeliness of cloud observation, so that the foundation cloud image classification is of great significance.

In recent decades, methods for classifying foundation cloud charts have been extensively studied. Traditional cloud classification methods rely on expert experience, are unreliable, time-consuming, and to some extent, rely on the experience of operators, with some uncertainty and bias in the classification results; in addition, human eye observation has gradually trended towards high costs.

The foundation cloud picture is a new natural texture image, which has attracted great attention in the field of computer vision in recent years, and the application of deep learning technology to the analysis and recognition research of the foundation cloud picture is increasing. The convolutional neural network technology is applied to cloud identification of a foundation cloud image, so that complex preprocessing of the cloud image in the early stage of image processing is avoided, the local receptive field of the convolutional neural network enables each neuron not to sense the whole image, only local sensing is needed, and all sensed information is integrated in the deep layer of the network to obtain the global information of the image; the weight sharing strategy of the method better accords with the characteristics of a biological neural network, greatly reduces the number of weight parameters, and reduces the computational complexity of the whole image processing process. The traditional convolutional neural network has high recognition rate in cloud classification, but has poor robustness in the cloud classification with occlusion.

Disclosure of Invention

In order to overcome the defects and shortcomings of the prior art, the invention provides a foundation cloud robust classification method based on weighted sparse representation and convolutional neural network feature fusion.

The technical scheme adopted by the invention is as follows:

a foundation cloud robust classification method comprises two parts, wherein the first part is used for feature extraction and converts an image into a feature vector y; the second part is used for classifying the input feature vector y; the method specifically comprises the following steps:

(1) the training samples are respectively processed by two convolutional neural networks to obtain the characteristic y₁∈R^n1×1,y₂∈R^n2×1N1 represents the dimension of the features obtained by the first convolutional neural network, n2 represents the dimension of the features obtained by the second convolutional neural network, R represents a real number space, and the two feature vectors are added to obtain a total feature vector y:

(2) let n1+ n2 be 2n, convert the 2 n-dimensional eigenvector into an n-dimensional eigenvector, and propose a projection composed of two projections P_iAnd P_eThe system comprises the following components:

denotes y passes through P_iVector after projection, z represents

Through P_eA projected vector;

two of which project P_iAnd P_eFrom training samples Y ═ Y₁ … Y_k … Y_K]Determining;

m_kindicates that the kth class training sample contains m_kPicture, K is 1,2, …, K;

(3) computing

Mean value v of_kSum variance vector

v_k，

i＝1,2,…,2n；

Y_k(j) represents a matrix Y_kColumn j of (1);

definition of

Let { V_i(j_p) Become V_i1.5n minimumSet of items, j_p＜j_p+1P is 1,2, …,1.5n-1, and then gives

Thus, it is possible to provide

k＝1,2,…,K；

Represents Y_kThrough projection P_iThe latter matrix;

(4) calculating V_k∈R^2n×KMean value v of^*Sum variance vector

1,2, …,1.5 n; let

Become into

N sets of maximum terms j_p＜j_p+1P is 1,2, …, n-1, then obtaining

(5) For all m_kOf a training sampleDictionary D of class k_kExpressed as:

D_k＝P_e(P_i(Y_k))；

k＝1,2,…,m_K；

and the whole dictionary D belongs to R^n×m，

Called extended dictionary, consisting of { D_kThe components are as follows:

D＝[D₁ … D_k … D_K]；

(6) the robust classification formula based on the optimal sparse representation is as follows:

a represents a sparse coefficient and a represents,

expressing the optimized sparse coefficient, and expressing a regular coefficient by lambda;

obtaining an optimal weighting matrix by iteration

(7) By obtaining

And

calculating delta_k：

δ_kRepresenting the weighted error between the test picture and each class;

then the

g_kRepresenting a weighted distance between the test picture and each class;

(8) the final test sample z is classified by the following formula:

preferably, in the step (6),

is estimated previously

Is updated specifically as follows:

the first iteration obtains the deviation

The first iteration obtains a weight matrix W^(l)：

Beta represents a decrement rate coefficient, phi represents a coefficient for controlling the position of a demarcation point;

followed by

The updating is as follows:

||·||_Frepresenting the F norm.

The invention has the beneficial effects that:

the invention relates to a weighted sparse representation-based convolutional neural network feature fusion cloud classification method, which is characterized in that two convolutional neural networks (inclusion-v 3 and ResNet-50) are extracted to serve as a weighted sparse representation dictionary to improve the operation efficiency; by weighting sparse representation classification, the robustness of the system can be improved under the condition of occlusion, and the performance better than that of a single convolutional neural network can be obtained by fusing two convolutional neural networks.

Drawings

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a test chart without noise (0%);

fig. 3 is a test chart after random noise (5% -25%) is added).

Detailed Description

The technical solutions of the present invention are further specifically described below by examples, which are for illustration of the present invention and are not intended to limit the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

Referring to fig. 1, a foundation cloud robust classification method includes two parts, the first part is used for feature extraction and converts an image into a feature vector y; the second part is used for classifying the input feature vector y; the method specifically comprises the following steps:

(1) the training samples are respectively processed by two convolutional neural networks (inclusion-v 3, ResNet-50) to obtain the characteristic y₁∈R^n1×1,y₂∈R^n2×1N1 represents the dimension of the features obtained by the first convolutional neural network, n2 represents the dimension of the features obtained by the second convolutional neural network, R represents a real number space, and the two feature vectors are added to obtain a total feature vector y:

(2) let n1+ n2 be 2n, convert the 2 n-dimensional eigenvector into an n-dimensional eigenvector, and propose a projection composed of two projections P_iAnd P_eThe system comprises the following components:

denotes y passes through P_iVector after projection, z represents

Through P_eA projected vector;

two of which project P_iAnd P_eFrom training samples Y ═ Y₁ … Y_k … Y_K]Determining;

m_kindicates that the kth class training sample contains m_kPicture, K is 1,2, …, K;

(3) computing

Mean value v of_kSum variance vector

v_k，

i＝1,2,…,2n；

Y_k(j) represents a matrix Y_kColumn j of (1);

definition of

Let { V_i(j_p) Become V_i1.5n sets of min terms, j_p＜j_p+1P is 1,2, …,1.5n-1, and then gives

Thus, it is possible to provide

k＝1,2,…,K；

Represents Y_kThrough projection P_iThe latter matrix;

(4) calculating V_k∈R^2n×KMean value v of^*Sum variance vector

i＝1,2,…,1.5n；Let

Become into

N sets of maximum terms j_p＜j_p+1P is 1,2, …, n-1, then obtaining

(5) For all m_kDictionary D of class k of training samples_kExpressed as:

D_k＝P_e(P_i(Y_k))；

k＝1,2,…,m_K；

and the whole dictionary D belongs to R^n×m，

Called extended dictionary, consisting of { D_kThe components are as follows:

D＝[D₁ … D_k … D_K]；

(6) the robust classification formula based on the optimal sparse representation is as follows:

a represents a sparse coefficient and a represents,

expressing the optimized sparse coefficient, and expressing a regular coefficient by lambda;

obtaining an optimal weighting matrix by iteration

Is estimated previously

Is updated specifically as follows:

the first iteration obtains the deviation

The first iteration obtains a weight matrix W^(l)：

Beta represents a decrement rate coefficient, phi represents a coefficient for controlling the position of a demarcation point;

followed by

The updating is as follows:

||·||_Frepresents the F norm;

(7) by obtaining

And

calculating delta_k：

δ_kRepresenting the error between the test picture and each class;

then the

g_kRepresenting the distance between the test picture and each class;

(8) the final test sample z is classified by the following formula:

TABLE 1

	0％	5％	10％	15％	20％	25％
							Inception-v3	96.97	84.03	83.18	82.24	79.43	77.52
ResNet-50	97.09	90.47	89.68	83.99	78.77	75.15
							The method of the invention	99.81	99.37	98.87	98.06	95.53	90.28

In order to verify the effectiveness and robustness of the method provided by the invention, experiments are carried out on MGCD data sets, 7 types of pictures are provided, the tested pictures are added with a plurality of random shielding noises, and the sizes of the pictures are 1024 x 1024. Comparing the recognition rates of the neural network and the method of the invention under the condition of adding random noise (0% -25%), verifying the effectiveness of the method, wherein the test image without noise (0%) is shown in figure 2, and the test image after adding random noise (5% -25%) is shown in figure 3; the experimental results are shown in table 1, and it can be seen from table 1 that, under the condition of not adding noise, the recognition rates of the method and the neural network are not much different, but the recognition rate of the neural network is rapidly reduced with the increase of noise, while the recognition rate of the method is slowly reduced, which indicates that the robustness of the method is stronger than that of a deep neural network.

Claims (2)

1. A foundation cloud robust classification method is characterized by comprising two parts, wherein the first part is used for feature extraction and converts an image into a feature vector y; the second part is used for classifying the input feature vector y; the method specifically comprises the following steps:

(1) the training samples are respectively processed by two convolutional neural networks to obtain the characteristic y₁∈R^n1×1,y₂∈R^n2×1N1 represents the dimension of the features obtained by the first convolutional neural network, n2 represents the dimension of the features obtained by the second convolutional neural network, R represents a real number space, and the two feature vectors are added to obtain a total feature vector y:

(2) let n1+ n2 be 2n, convert the 2 n-dimensional eigenvector into an n-dimensional eigenvector, and propose a projection composed of two projections P_iAnd P_eThe system comprises the following components:

denotes y passes through P_iVector after projection, z represents

Through P_eA projected vector;

two of which project P_iAnd P_eFrom training samples Y ═ Y₁…Y_k…Y_K]Determining;

m_krepresenting class k training samplesContaining m_kPicture, K is 1,2, …, K;

(3) computing

Mean value v of_kSum variance vector

v_k，

i＝1,2,…,2n；

Y_k(j) represents a matrix Y_kColumn j of (1);

definition of

Let { V_i(j_p) Become V_i1.5n sets of min terms, j_p＜j_p+1P is 1,2, …,1.5n-1, and then gives

Thus, it is possible to provide

k＝1,2,…,K；

Represents Y_kThrough projection P_iThe latter matrix;

(4) calculating V_k∈R^2n×KMean value v of^*Sum variance vector

Let

Become into

1,2, …,1.5 n;

j_p＜j_p+1p is 1,2, …, n-1, then obtaining

(5) For all m_kDictionary D of class k of training samples_kExpressed as:

D_k＝P_e(P_i(Y_k))；

k＝1,2,…,m_K；

and the whole dictionary D belongs to R^n×m，

Called extended dictionary, consisting of { D_kThe components are as follows:

D＝[D₁…D_k…D_K]；

(6) the robust classification formula based on the optimal sparse representation is as follows:

a represents a sparse coefficient and a represents,

expressing the optimized sparse coefficient, and expressing a regular coefficient by lambda;

obtaining an optimal weighting matrix by iteration

(7) By obtaining

And

calculating delta_k：

δ_kRepresenting the weighted error between the test picture and each class;

then the

g_k＝||δ_k||₂,

g_kRepresenting a weighted distance between the test picture and each class;

(8) the final test sample z is classified by the following formula:

2. the ground-based cloud robust classification method according to claim 1, wherein in step (6),

is estimated previously

Is updated specifically as follows:

the first iteration obtains the deviation

The first iteration obtains a weight matrix W^(l)：

Beta represents a decrement rate coefficient, phi represents a coefficient for controlling the position of a demarcation point;

followed by

The updating is as follows:

||·||_Frepresenting the F norm.

Images