CN108805039A - The Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features - Google Patents

Abstract

The invention belongs to the technical field of radar radiation source signal modulation recognition, and in particular relates to a modulation recognition method for extracting time-frequency image features by combining entropy and pre-trained CNN. Firstly, time-frequency transformation is performed on the 9 types of radar signal sets to be identified to obtain a time-frequency image; then, based on the pre-trained convolutional neural network model imagenet‑vgg‑verydeep‑19 provided by MatConvNet official website, the input layer is connected to the fc6 fully connected layer Constitute the FT-VGGNet-fc6 feature migration extraction module; then send the adjusted image into the feature migration extraction module, and output the radar signal time-frequency image features; then grayscale the adjusted image, and manually extract the Renyi entropy of the processed image; Next, divide the training set and test set according to a certain ratio, and select the training set to train the SVM classifier; finally, use the trained SVM classifier to identify the training set of time-frequency images, and use the multi-signal-to-noise ratio under 9 categories A dataset composed of radar signals validates the recognition rate of the FT‑VGGNET‑fc6‑SVM classifier.

Description

Modulation recognition method for extracting time-frequency image features by joint entropy and pre-trained CNN

technical field

The invention belongs to the technical field of radar radiation source signal modulation recognition, and in particular relates to a modulation recognition method for extracting time-frequency image features by combining entropy and pre-trained CNN.

Background technique

Radar emitter signal modulation identification is an important link in electronic countermeasures and electronic reconnaissance, and plays a very important role in electronic warfare. The commonly used radar radiation source signal modulation recognition methods include the signal modulation recognition method based on five parameters, the intrapulse modulation recognition method based on time-frequency images, and the signal modulation recognition method based on CNN.

The traditional modulation recognition method based on five parameters (carrier frequency, pulse arrival time, pulse amplitude, pulse width, and pulse arrival direction) can obtain other characteristic parameters of the pulse after multiple measurements, but it does not consider the pulse of the new system radar. Due to the internal modulation characteristics, effective identification cannot be realized. At the same time, this method relies too much on the original information in the database, and cannot identify radars that do not exist in the database and cannot self-learn, so it has little effect in front of new system radars.

The intra-pulse modulation recognition method based on time-frequency images is widely used. It can convert radar signals into time-frequency images through time-frequency distribution, extract features after image preprocessing, and access support vector machine (SVM) at its back end. ) classifier, even at a low signal-to-noise ratio can obtain a considerable recognition rate. However, this method requires manual extraction of image features, and the speed of manual feature extraction is slow. In addition, improper manual extraction of feature information will inevitably lead to deviations in recognition, which will eventually lead to a low recognition rate.

Compared with the above two methods, the CNN-based signal modulation recognition method can automatically complete the extraction of time-frequency image features when the signal modulation recognition method is used to recognize the radar radiation source signal, and realize a high degree of automation of feature extraction. However, after using CNN to extract features, there are too many output features, and redundant information will reduce the effectiveness of the system, and the extracted features have poor recognition effect under low signal-to-noise ratio conditions, and the corresponding system has weak anti-noise ability. At the same time, when faced with small sample data, the sharp drop in the number of neural units between fully connected layers of CNN will lead to an overall decline in the system's recognition rate.

Contents of the invention

The purpose of the present invention is to propose a modulation recognition method that combines entropy and pre-trained CNN to extract time-frequency image features, utilizes CNN to automatically extract image features to realize automatic feature extraction, and applies principal component analysis (PCA) to reduce the output feature dimension and improve the system Effectiveness, combining the features after dimensionality reduction with the artificially extracted Renyi entropy to improve the recognition rate of the system under the condition of low signal-to-noise ratio, and applying SVM to solve the problem of low training accuracy of small samples of deep networks, and finally realize accurate and fast detection of radar signals identify.

The purpose of the present invention is achieved like this:

Joint entropy and pre-training CNN extract the modulation recognition method of time-frequency image feature, it is characterized in that, comprises the following steps:

Step 1 generates a radar signal set composed of CW, LFM, BPSK, COSTAS, FRANK, P1, P2, P3, and P4 according to nine types of radar signal parameters; the nine types of radar signal parameters include: sampling frequency, number of sampling points, CW modulation carrier frequency, Barker code, bandwidth, frequency sequence, starting frequency, carrier frequency generated by P3 and P4 modulation methods, and carrier frequencies generated by FRANK, P1, P2, and modulation methods;

Step 2 Carry out time-frequency analysis on the radar signal, and apply CWD to perform time-frequency transformation on the radar signal set to be identified to obtain a time-frequency image; the corresponding CWD formula is as follows:

The time-varying local correlation function can be obtained by performing sliding window processing on the correlation function:

When the window function takes the time impact function, no restriction is imposed, and the instantaneous value is taken in the time domain:

The Wigner Ville distribution (WVD) can be obtained by Fourier transforming the time-varying local correlation function:

WVD can be added by adding kernel function Get CWD:

Step 3 The pre-training convolutional neural network model uses imagenet-vgg-verydeep-19 provided by the official website of MatConvNet, and converts the image into a 224×224×3 image; keep the network parameters of imagenet-vgg-verydeep-19 unchanged, by Its Input input layer to fc6 fully connected layer constitutes the FT-VGGNet-fc6 feature migration extraction module;

Step 4 Send the image set adjusted in step 3 to the feature migration extraction module to generate radar signal time-frequency image features, and use the FT-VGGNet-fc6 feature migration extraction module retained to the fc6 fully connected layer to obtain 8×512 radar signals Time-frequency image features;

Step 5. The FT-VGGNet-fc6 feature migration and extraction module outputs too many features, and redundant information will appear, resulting in a decrease in training speed and system effectiveness. Use PCA dimensionality reduction to retain 95 features with significant discrimination;

PCA uses the time-frequency image as the original sample to form a data matrix:

Its covariance matrix is R=XX ^T , and the eigenvalue decomposition of the covariance matrix can be performed:

R _M×M ＝U∧U ^T

In the formula, T represents the transpose, ∧ is the eigenvalue diagonal matrix of the covariance matrix, U is the corresponding eigenmatrix, and the time-frequency image is transformed as follows:

P _M×N ＝U ^T X＝[p ₁ ,p ₂ ,…,p _M ] ^T

In the formula, P is the principal component of the time-frequency image binary matrix, p ₁ is the first principal component, p _j is the jth principal component, and the first k principal components are selected to form the feature matrix of the time-frequency image;

Step 6 Perform image grayscale on the adjusted time-frequency image, and the brightness of the color image pixel corresponding to the point with three components of R, G, and B can be calculated by the grayscale formula as follows:

I=0.3B+0.59G+0.11R

Step 7 Extract the Renyi entropy in the grayscale image that can improve the recognition rate of the system, especially under the condition of low signal-to-noise ratio, and make it normalized and combined with the 95 features after dimensionality reduction to form 100 new radar signal time-frequency Image features; the Renyi entropy of the time-frequency image is expressed as:

In the formula, P ^α (t,f) represents the time-frequency distribution of the signal; for the selection of the Renyi entropy order α, regardless of the case that the non-integer order α produces complex entropy values, the selected order is 3, 5, 7, The Renyi entropy of 9 and 11 is used as the identification feature of the signal;

Step 8: For nine kinds of radar signals, select 300 time-frequency images for each type of radar signal when the signal-to-noise ratio is 0dB, and randomly divide the time-frequency image features of radar signals into a training set and a test set in a ratio of 7:3;

Step 9 Divide the training set and the test set, combine the particle swarm optimization algorithm (PSO) and the artificial bee colony algorithm (ABC), and jointly apply the two algorithms to calculate the fitness value. Each independent calculation is compared and the best is taken until the maximum number of iterations is reached. The optimal value realizes the joint optimization of the SVM parameters by the two algorithms, and selects the training set to train the SVM classifier;

The search behavior of PSO particles in the particle swarm algorithm is affected by the search behavior of other particles in the group, and it is easy to fall into the local optimal solution, while the artificial bee colony algorithm ABC has weak local search ability and slow convergence speed; Local solution, the combination of the two improves the optimization ability of the PSO algorithm, and the related formula is as follows:

The update of the particle position and velocity in the particle swarm is:

The search equation in artificial bee colony algorithm is:

v _ij ＝x _ij +φ _ij (x _ij -x _kj )

First, the population size, PSO speed range, ABC nectar source number and parameters, (C, σ) and initial fitness value of ABC algorithm and PSO algorithm should be initialized; then divided into two independent subpopulations and jointly applied ABC algorithm and PSO algorithm Calculate the fitness value, and take the best for each independent calculation; finally, the optimal fitness value when the two reach the maximum number of iterations is used as the optimal parameter of SVM, and realize the joint optimization of the parameters of SVM by applying the two algorithms;

Input the features of the training set and the corresponding category labels into the above SVM parameter optimization algorithm for cross-validation, determine the σ and C of the SVM, and use the training set to train the SVM according to the best parameters;

Step 10: Select a total of 32,400 time-frequency images of 9 types of radar signals with a signal-to-noise ratio of -3 to 8dB, among which, 300 time-frequency images of each type of radar signal with a single signal-to-noise ratio; Randomly select 210 images of each type of radar signal with a single signal-to-noise ratio, a total of 22,680 images as a training set, and the remaining 90 images of each type of radar signal with a single signal-to-noise ratio, a total of 9,720 images are used as a test set; use the trained SVM classification The machine recognizes the training set of time-frequency images, and verifies the recognition rate of the FT-VGGNET-fc6-SVM classifier.

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

1. It is proposed to combine Renyi entropy with CNN to extract time-frequency image features, access SVM and combine transfer learning theory to modulate and recognize radar signals, which overcomes the respective limitations of CNN and SVM. Compared with the current recognition algorithm, its recognition speed Faster, higher recognition rate (especially at low signal-to-noise ratio), better effectiveness and higher robustness of the recognition system;

2. As a deep network, CNN can obtain more representative information, so that the modulation recognition features are more comprehensive and effective. The application of CNN to automatically extract image features solves the problems of slow and incomplete extraction of features manually, and can improve the performance of the system. stability;

3. The FT-VGGNet-fc6 feature migration module constructed by image input, CNN extracts large-scale output features, and the application of PCA dimensionality reduction can effectively find out the most important 95 features, remove redundant and invalid information, and make the recognition system more efficient. The effectiveness far exceeds the system that simply applies CNN modulation recognition;

4. Renyi entropy has the strong anti-noise ability common to entropy features, especially at low signal-to-noise ratios, the recognition rate is higher than that based on other features. Therefore, Renyi entropy in preprocessed images is manually extracted after CNN extracts feature dimensionality reduction. Co-normalization makes the combination of the two improve the noise immunity of the recognition system;

5. Using SVM requires a very small amount of training samples to obtain the characteristics of the category pattern, which effectively avoids the problem of a steep drop in the number of neural units between fully connected layers of the convolutional neural network due to too few sample categories, and the application of SVM can make The generalization performance and robustness of the system are effectively improved;

6. The set FT-VGGNET-fc6-SVM can be applied to the SVM classifier training when facing a small sample, and when facing a large amount of data, the feature extraction parameters can be fixed for feature migration, so that the radar signal data set can use large The database trains the entire network to obtain more discriminative and robust features.

Description of drawings

Fig. 1 is the imagenet-vgg-verydeep-19 network structural diagram that the present invention relates to;

Fig. 2 is the modulation recognition flow chart of joint entropy and pre-training CNN extraction time-frequency image feature of the present invention;

Fig. 3 is a test result graph of the FT-VGGNET-fc6-SVM classifier of the present invention.

Detailed ways

The present invention will be further described below in conjunction with accompanying drawing:

First, the Choi-Williams distribution (CWD) time-frequency transformation is performed on the 9 types of radar signal sets to be identified to obtain a time-frequency image; then based on the pre-trained convolutional neural network model imagenet-vgg-verydeep-19 provided by MatConvNet official website, the time-frequency The image size is adjusted to 224×224×3, so that the parameters of the pre-trained network model remain unchanged, and the FT-VGGNet-fc6 feature migration extraction module is composed of its Input input layer to the fc6 fully connected layer; then the adjusted image is sent to the feature The migration extraction module outputs the time-frequency image features of the radar signal, and applies PCA to reduce the feature dimension to 95 features; then grayscales the adjusted image, manually extracts the Renyi entropy of the processed image, and combines the dimensionality reduction feature with the Renyi entropy After normalization, a new radar signal time-frequency image feature is formed; next, the training set and the test set are divided according to a certain ratio, and the particle swarm optimization algorithm (PSO) and the artificial bee colony algorithm (ABC) are combined, and the two algorithms are jointly applied to calculate the adaptive Value, each independent calculation is compared to take the optimal value, until the maximum number of iterations is reached to take the optimal value, realize the joint optimization of the SVM parameters by the two algorithms, and select the training set to train the SVM classifier; finally, use the trained SVM to classify The FT-VGGNET-fc6-SVM classifier is used to identify the training set of time-frequency images, and the recognition rate of the FT-VGGNET-fc6-SVM classifier is verified by using a data set composed of 9 types of radar signals under multiple signal-to-noise ratios.

Specifically include the following steps:

Step 1: Generate a radar signal set consisting of CW, LFM, BPSK, COSTAS, FRANK, P1, P2, P3, and P4 according to the 9 types of radar signal parameter values provided in Table 1.

Table 1

Step 2: The analysis and processing of non-stationary signals such as radar signals cannot be limited to the time domain or frequency domain only. Time-frequency analysis should be performed on radar signals, and CWD can effectively suppress while obtaining better time and frequency resolution. Intersection term, applying CWD to time-frequency transform the radar signal set to be identified to obtain a time-frequency image, the corresponding CWD formula is as follows:

The time-varying local correlation function can be obtained by performing sliding window processing on the correlation function:

When the window function takes the time impact function, no restriction is imposed, and the instantaneous value is taken in the time domain:

The Wigner Ville distribution (WVD) can be obtained by Fourier transforming the time-varying local correlation function:

WVD can be added by adding kernel function Get CWD:

Step 3: The pre-training convolutional neural network model uses imagenet-vgg-verydeep-19 provided by MatConvNet official website. Since the imagenet-vgg-verydeep-19 model needs to input a fixed-size three-primary-color (RGB) image, it is necessary to convert the image to 224 ×224×3 size image. Keep the imagenet-vgg-verydeep-19 network parameters unchanged, and form the FT-VGGNet-fc6 feature migration extraction module from its Input input layer to the fc6 fully connected layer.

Step 4: Send the image set adjusted in Step 3 into the feature migration extraction module to generate radar signal time-frequency image features, and use the FT-VGGNet-fc6 feature migration extraction module reserved to the fc6 fully connected layer to obtain

8×512 radar signal time-frequency image features.

Step 5: The FT-VGGNet-fc6 feature migration and extraction module outputs too many features, and redundant information will appear, resulting in a decrease in training speed and system effectiveness. PCA dimensionality reduction can be used to retain 95 features with significant discrimination. Specifically, PCA takes the time-frequency image as the original sample to form a data matrix:

Its covariance matrix is R=XX ^T , and the eigenvalue decomposition of the covariance matrix can be performed:

R _{M × M} = U∧U ^T (6)

In the formula, T represents the transpose, ∧ is the eigenvalue diagonal matrix of the covariance matrix, U is the corresponding eigenmatrix, and the time-frequency image is transformed as follows:

P _M×N ＝U ^T X＝[p ₁ ,p ₂ ,…,p _M ] ^T (7)

In the formula, P is the principal component of the time-frequency image binary matrix, p ₁ is the first principal component, p _j is the jth principal component, and the first k principal components are selected to form the feature matrix of the time-frequency image.

Step 6: Perform image grayscale on the adjusted time-frequency image, and the brightness of the color image pixel corresponding to the point with three components of R, G, and B can be calculated by the grayscale formula as follows:

I＝0.3B+0.59G+0.11R (8)

Step 7: Extract the Renyi entropy in the grayscale image that can improve the recognition rate of the system, especially under the condition of low signal-to-noise ratio, and make it normalized and combined with the 95 features after dimensionality reduction to form 100 new radar signals frequency image features. The Renyi entropy of the time-frequency image is expressed as:

In the formula, P ^α (t, f) represents the time-frequency distribution of the signal. For the selection of the Renyi entropy order α, this embodiment does not consider the case that the non-integer order α produces complex entropy values, and mainly selects the Renyi entropy with orders 3, 5, 7, 9, and 11 as the identification feature of the signal.

Step 8: For 9 kinds of radar signals, select 300 time-frequency images for each type of radar signal when the signal-to-noise ratio is 0dB, and randomly divide the time-frequency image features of the radar signals into a training set and a test set in a ratio of 7:3.

Step 9: Particle swarm optimization (PSO) particle search behavior is affected by the search behavior of other particles in the group and easily falls into a local optimal solution, while artificial bee colony algorithm (ABC) has weak local search ability and slow convergence speed. The ABC algorithm can make up for the PSO algorithm falling into a local solution. The combination of the two can improve the optimization ability of the PSO algorithm. The related formula is as follows:

The update of the particle position and velocity in the particle swarm is:

The search equation in artificial bee colony algorithm is:

v _ij ＝x _ij +φ _ij (x _ij -x _kj ) (11)

First, the population size, PSO speed range, ABC nectar source number and parameters, (C, σ) and initial fitness value of ABC algorithm and PSO algorithm should be initialized; then divided into two independent subpopulations and jointly applied ABC algorithm and PSO algorithm The fitness value is calculated, and the optimal value is selected for each independent calculation; finally, the optimal fitness value when the two reach the maximum number of iterations is used as the optimal parameter of the SVM, and the joint optimization of the SVM parameters is realized by applying the two algorithms.

The characteristics of the training set combined with the corresponding category labels are input into the above SVM parameter optimization algorithm for cross-validation, the σ and C of the SVM are determined, and the SVM is trained using the training set according to the best parameters.

Step 10: Select a total of 32,400 time-frequency images of 9 types of radar signals with a signal-to-noise ratio of -3 to 8dB, among which, 300 time-frequency images of each type of radar signal with a single signal-to-noise ratio. Randomly select 210 images from each type of radar signal with a single signal-to-noise ratio in the data set, a total of 22,680 images as a training set, and the remaining 90 images with a single signal-to-noise ratio for each type of radar signal in the data set, a total of 9,720 images as a test set. Use the trained SVM classifier to recognize the training set of time-frequency images, and verify the recognition rate of the FT-VGGNET-fc6-SVM classifier.

Figure 3 shows the test results of the FT-VGGNet-fc6-SVM classifier, and its recognition rate increases with the increase of the signal-to-noise ratio, and tends to be smooth after 5dB. A recognition rate of 87% was obtained at -3dB, and then it showed a steady upward trend, and reached more than 99% at 7dB, indicating that the present invention has a higher recognition rate (especially at low signal-to-noise ratio) for radar emitter signal modulation. In addition, the test of the present invention is stable, and the recognition rate does not fluctuate greatly, which shows that the present invention has good stability and robustness, and has good practicability.

The present invention provides a modulation recognition method that combines entropy and pre-trained CNN to extract time-frequency image features. There are many methods and approaches to realize this technical solution, and the above is only a preferred embodiment of the present invention. All components that are not specified in this embodiment can be realized by existing technologies.

Claims (1)

1. the Modulation Identification method of combination entropy and pre-training CNN extraction time-frequency image features, which is characterized in that including following step Suddenly：

Step 1 is made of according to nine class radar signal parameters, generation CW, LFM, BPSK, COSTAS, FRANK, P1, P2, P3, P4 Radar signal collection；The nine class radar signal parameters include：Sample frequency, sampling number, CW modulation systems carrier frequency, Bark Code, bandwidth, frequency sequence, initial frequency, the carrier frequency of P3 and P4 modulation systems generation and FRANK, P1, P2, modulation system production Raw carrier frequency；

Step 2 carries out time frequency analysis to radar signal, and radar signal collection to be identified, which is carried out time-frequency conversion, using CWD obtains Time-frequency image；Corresponding CWD formula are as follows：

The local correlation function of time-varying is obtained by pair correlation function as slide window processing：

It is without restriction when window function takes time impulse function, and take instantaneous value in time domain：

Clock synchronization becomes local correlation function and makees Fourier transformation, obtains Wigner Ville distributions (WVD)：

WVD is by adding kernel functionObtain CWD：

The imagenet-vgg- that the official websites step 3 pre-training convolutional neural networks model selection MatConvNet provide Picture is converted to the image of 224 × 224 × 3 sizes by verydeep-19；Make imagenet-vgg-verydeep-19 networks Parameter remains unchanged, and constituting FT-VGGNet-fc6 features by its Input input layer to the full articulamentums of fc6 migrates extraction module；

Image set after step 4 adjusts step 3 is sent into feature migration extraction module, and it is special to generate radar signal time-frequency image Sign obtains 8 × 512 radar signal time-frequencies using the FT-VGGNet-fc6 features migration extraction module retained to the full articulamentums of fc6 Characteristics of image；

The feature that step 5 FT-VGGNet-fc6 features migrate extraction module output is excessive, has redundancy and occurs causing to instruct Practice speed and system effectiveness declines, retains 95 features with notable discrimination with PCA dimensionality reductions；

Time-frequency image is constituted a data matrix by PCA：

Its covariance matrix is R=XX^T, Eigenvalues Decomposition is made to the covariance matrix：

R_M×M=U ∧ U^T

In formula, T indicates that transposition, ∧ are the characteristic value diagonal matrix of covariance matrix, and U is corresponding eigenmatrix, to time-frequency image Make such as down conversion：

P_M×N=U^TX=[p₁,p₂,…,p_M]^T

In formula, P is the principal component of time-frequency image two values matrix, p₁It is first principal component, p_jFor jth principal component, k master before choosing Ingredient constitutes the eigenmatrix of time-frequency image；

Step 6 carries out image gray processing to the time-frequency image after adjustment, and the color image pixel of three-component R, G and B correspond to should The brightness of point is calculated as with gray processing formula：

I=0.3B+0.59G+0.11R；

The Renyi entropys of discrimination under system recognition rate especially Low SNR can be improved in step 7 extraction gray level image, and It is allowed to constitute 100 new radar signal time-frequency image features with common normalized combine of 95 features after dimensionality reduction；Time-frequency figure The Renyi entropys of picture are expressed as：

In formula, P^α(t, f) indicates the time-frequency distributions of signal；Selection for Renyi entropy exponent numbers α does not consider that non-integral order α is generated The case where entropy of plural number, chooses identification feature of the Renyi entropys that exponent number is 3,5,7,9,11 as signal；

Step 8 is directed to nine kinds of radar signals, chooses every class radar signal each 300 of time-frequency image when signal-to-noise ratio is 0dB, will Radar signal time-frequency image feature presses 7:3 ratio is randomly divided into training set and test set；

Step 9 divides training set and test set, and particle cluster algorithm (PSO) and artificial bee colony algorithm (ABC) are combined, combines and answers Adaptive value is calculated with two algorithms, once relatively takes optimal per independent calculate, takes optimal value until reaching maximum iteration, realization Two algorithms are combined to SVM parameter optimizations, and are chosen training set and be trained to SVM classifier；

The search behavior of particle cluster algorithm PSO particles is influenced by the search behavior of other particles in group and is easily absorbed in part most Excellent solution, and artificial bee colony algorithm ABC local search abilities are weaker and convergence rate is slow；ABC algorithms can overcome the disadvantages that PSO algorithms are fallen into Enter local solution, the combination of the two improves the optimizing ability of PSO algorithms, and correlation formula is as follows：

Particle position and speed is updated in population：

Equation is searched in artificial bee colony algorithm is：

v_ij=x_ij+φ_ij(x_ij-x_kj)

First having to will be (C, the σ) of population scale, PSO velocity intervals, the nectar sources ABC number and parameter, ABC algorithms and PSO algorithms and first Beginning adaptive value is initialized；Then 2 sub- populations of independence are divided into and use in conjunction ABC algorithms and PSO algorithms are calculated and adapted to Value once relatively takes optimal per independent calculating；Adaptive optimal control value when the two finally to reach to maximum iteration as SVM most Excellent parameter realizes the parameter optimization combined to SVM using two algorithms；

Training set feature combination respective classes label is inputted into above-mentioned SVM parameter optimizations algorithm and carries out cross validation, determines SVM's σ and C is trained SVM using training set according to optimal parameter；

Step 10：Choose 9 class radar signal time-frequency images totally 32400 pictures in the case of signal-to-noise ratio is -3~8dB, wherein Per each 300 of time-frequency image under the single signal-to-noise ratio of class radar signal；Under every single signal-to-noise ratio of class radar signal of data set Select 210 at random, totally 22680 are used as training set, data set per class radar signal single signal-to-noise ratio under remaining 90, altogether 9720 are used as test set；The training set of time-frequency image is identified using the SVM classifier after training, and verifies FT- The discrimination of VGGNET-fc6-SVM graders.