CN109033978A - A kind of CNN-SVM mixed model gesture identification method based on error correction strategies - Google Patents

Abstract

The invention provides a CNN-SVM hybrid model gesture recognition method based on an error correction strategy, which belongs to the field of human-computer interaction. The CNN-SVM hybrid model gesture recognition method based on the error correction strategy first preprocesses the collected gesture data, then automatically extracts features and performs prediction classification to obtain classification results, and finally uses the error correction strategy to correct the classification results . By using the method of the invention, the misrecognition rate between easily confused gestures is reduced, and the recognition rate of static gestures is improved.

Description

A CNN-SVM Hybrid Model Gesture Recognition Method Based on Error Correction Strategy

technical field

The invention belongs to the field of human-computer interaction, and in particular relates to a CNN-SVM hybrid model gesture recognition method based on an error correction strategy.

Background technique

As computers become more and more popular in today's society, a convenient and natural human-computer interaction (HCI) method is particularly important for users. Among the many human-computer interaction methods, gestures, as a natural, concise and intuitive human-computer interaction method, have attracted more and more people's attention, and they can play an important role in various real-world scenarios, such as somatosensory Games, sign language recognition, smart wearable devices and smart teaching. The purpose of gesture recognition is to design an algorithm to enable the computer to recognize the gestures of pictures or human bodies, understand the meaning of gestures, and realize the interaction between humans and computers. In the process of gesture recognition, gestures are usually in a complex environment. In order to accurately perform human-computer interaction, the designed gesture recognition algorithm should have good recognition ability in various light, angles, backgrounds and other complex environments.

Traditional gesture recognition algorithms are mainly based on Hidden Markov Model (HMM) and template matching. Among them, the gesture recognition method based on the hidden Markov model can be used to express a Markov process with hidden unknown parameters, and the process of gesture recognition can be regarded as a Markov chain containing time series, so This model can be applied to gesture recognition. The gesture recognition method based on last-shift matching takes the contour, edge, and spatial distribution of the gesture as features to establish a gesture template, and applies the template matching algorithm to realize gesture recognition. These two methods require manual extraction of features, and manual extraction of gesture features requires a large amount of experience, and the manual extraction of features has certain subjectivity and limitations, making it easy to ignore some salient features, so traditional methods often identify Capacity is limited and not efficient.

Convolutional Neural Network (CNN) is one of the most widely used models in the field of machine vision and image processing at present. The convolutional neural network can obtain the local and global features of the input image through training and learning, which solves the problem of manual extraction of feature bands. The problem of insufficient feature extraction. In recent years, convolutional neural networks have been successfully applied to image retrieval, face recognition, expression recognition and object detection. Some scholars have applied CNN to the field of gesture recognition. Jawad Nagi et al. combined the maximum pooling layer and convolutional neural network (MPCNN) for gesture recognition and achieved good results. Takayoushi et al. proposed an end-to-end The deep convolutional network realizes gesture recognition and improves the accuracy of gesture recognition at the same time. In the application of gesture recognition, relatively shallow networks are generally used. In the traditional static gesture recognition method, the gesture recognition method based on manual feature extraction takes a long time and the recognition rate is low.

Contents of the invention

The purpose of the present invention is to solve the problems existing in the above-mentioned prior art, and provide a CNN-SVM hybrid model gesture recognition method based on an error correction strategy. Misrecognition rate of confusing gestures, and finally realize the recognition of static gestures.

The present invention is achieved through the following technical solutions:

A CNN-SVM hybrid model gesture recognition method based on an error correction strategy. First, preprocess the collected gesture data, then automatically extract features and perform prediction classification to obtain classification results, and finally use an error correction strategy to perform classification on the classification results. correct.

The methods include:

Step 1: Preprocess the collected data to obtain training samples and test samples;

The second step: obtain the CNN-SVM hybrid model;

The third step: input the test sample into the CNN-SVM hybrid model obtained in the second step, and obtain the classification result and the probability estimation of the classification result and the confusion matrix;

Step 4: Get the error correction strategy based on the probability estimate obtained in the third step and the confusion matrix, and then use the error correction strategy to correct the classification results.

The operation of the first step includes:

(11) Static gestures are collected, and depth images and color images of the hands are obtained respectively;

(12) Processing the depth image to obtain a mask image;

(13) performing an AND operation on the color image and the mask image to obtain a rough gesture area image;

(14) Using the Bayesian skin color model to perform skin color segmentation on the rough gesture region image to obtain a segmented image, the segmented image is divided into two parts, one part is used as a training sample, and the other part is used as a test sample.

In described step (11), adopt Kinect to gather static gesture.

Described second step is realized like this: replace the last output layer of CNN classifier with SVM classifier

The operations of the second step include:

(21) the training sample is input to the input layer of the CNN classifier, until the training process converges or reaches the maximum number of iterations through the training of the CNN classifier, obtains the trained CNN model;

(22): inputting the training sample into the trained CNN model to perform automatic feature extraction to obtain the feature vector of the training sample;

(23): Input the feature vector of the training sample into the SVM classifier for secondary training, and obtain the CNN-SVM hybrid model after the training is completed.

The error correction strategy refers to specifying a threshold, filtering out wrong classification results according to the threshold, and then correcting the final classification results according to statistical data obtained from experiments.

The operations of the fourth step include:

In the N classification problem, let M _i be a threshold for error correction of all test samples with classification result i, and the description of M _i is as follows:

Among them, M _{i, j} represents the mean value calculated by the sample whose predicted result is i, but the actual value is j, and M _i is a j-dimensional vector; S _{i, j} represents all the samples whose predicted result is i, but the true value is j The number of samples, S _i represents the number of all test samples predicted to be class i, P _n (i) represents the maximum value of the probability estimate of the nth test sample among all test samples predicted to be class i, P _n ( j) represents the second largest value; i represents the class to which the maximum value belongs in the classification estimation, and j represents the class to which the second largest value belongs to in the classification estimation;

When the probability estimate satisfies the following conditions, modify the class corresponding to the maximum value of the probability estimate to the class corresponding to the second largest value:

Among them, w _n (i) represents the distance between the maximum value of the probability estimate and the second maximum value of the probability estimate for the prediction result of class i, that is, it is equal to P _n (i)-P _n (j) in value, and p _ij is represented in the confusion matrix The probability that the classification result is i but the true value is j.

Compared with the prior art, the beneficial effect of the present invention is that: using the method of the present invention reduces the misrecognition rate between confusing gestures and improves the recognition rate of static gestures.

Description of drawings

Figure 1-1. Photographs of nine different gestures

Figure 1-2 Depth images corresponding to the nine different gestures in Figure 1-1

Fig. 2 block diagram of insufficient image preprocessing in the method of the present invention

Figure 3 Pictures during preprocessing

The CNN network structural diagram that Fig. 4 method of the present invention adopts

Figure 5 Curves of test accuracy on different data sets

Fig. 6 is a block diagram of the steps of the method of the present invention.

Detailed ways

Below in conjunction with accompanying drawing, the present invention is described in further detail:

Combining the advantages of convolutional neural network and support vector machine, the present invention proposes a hybrid model to automatically extract features and improve the generalization ability of the model, and uses an error correction strategy based on probability estimation to reduce the error of confusing gestures. literacy rate.

As shown in FIG. 6 , the method of the present invention includes: firstly, segmental preprocessing is performed on the gesture data collected by Kinect to reduce interference from complex backgrounds and other parts of the human body. The hybrid model then automatically extracts features and performs predictive classification. Finally, the classification decision is adjusted using an error correction strategy. Experiments are carried out on the established database, and finally the recognition rate without using error correction strategy is 95.81%, and the average accuracy rate is 97.32% after using error correction strategy.

Data collection in the inventive method is as follows:

The system of the present invention adopts Kinect2.0 to collect static gestures, acquires depth images and color images of hands respectively, and then establishes corresponding gesture databases. The established gesture library contains a total of 17 types of gestures, which are composed of static images collected by 300 college students under different lighting backgrounds. In the present invention, 9 kinds of gestures commonly used by human beings are selected, and each gesture contains 3300 pictures. Figure 1-2 and Figure 1-2 are photos and depth images of 9 gestures completed by the operator, respectively.

Data preprocessing is as follows:

It is not difficult to see from the collected gesture images that although the image of human gestures in the color image is clearly identifiable, it is still very difficult to complete accurate recognition, because the collected gestures are affected by the angle of view, appearance, shape, other parts of the human body and The effect of complex backgrounds. In the collected depth image, on the one hand, the depth information is not affected by the color, texture characteristics and illumination of the human hand itself, so it has good robustness and high precision; on the other hand, the depth information in the depth image reflects the human hand to collect The distance between the devices, so the depth difference in the gesture area is not very big. Because the depth image has been segmented during the acquisition process, using this feature can help segment the gesture area of interest in the color image, thereby reducing the interference of other parts of the human body and complex backgrounds in the color image. The steps of segmentation preprocessing are shown in Fig. 2.

In the preprocessing process, the present invention binarizes the collected depth image. Since the depth image has been converted into a grayscale depth image during the collection process, the value range of the depth value is adjusted to the grayscale value [0- 255] between. Since the gesture region has been segmented on the depth map during the acquisition process, the binary image of the gesture region can be obtained by using the gray value. Directly performing logical AND operations on the mask image (set a threshold value of 128 for the grayscale image in the present invention, the pixel points greater than 128 are assigned a value of 1, and the pixel points smaller than 128 are assigned a value of 0) and the color image can only get rough Gesture area images, due to the problem of different resolutions between the depth image and the color image during the Kinect acquisition process, there will be interference from non-gesture pixels around. Perform skin color segmentation on the obtained rough gesture area, using the Bayesian skin color model (please refer to the literature "M.J. Jones, et al. Statistical color models with application to skin detection [J]. International Journal of Computer Vision (IJCV), 2002, 46 (1):81-96") to obtain accurate gesture region images.

In the present invention, an image is randomly selected to test the effectiveness of segmentation preprocessing, in which the color image, depth image, mask image, rough gesture area, and segmented image are shown in Figure 3-1 to Figure 3-5 respectively .

It can be clearly seen that the segmentation preprocessing in the method of the present invention can effectively remove the influence of complex backgrounds and other parts of the human body, and finally, the use of the Bayesian skin color model can also accurately retain the effective information of the gesture area for later training work Provides good data protection.

The hybrid CNN-SVM model is as follows:

SVM classifier: The support vector machine converts the linearly inseparable samples of the low-dimensional input space into the high-dimensional feature space by selecting different kernel functions to make them linearly separable, and constructs the optimal model in the feature space based on the principle of institutional risk minimization. Hyperplane obtains a structured description of data distribution, thus reducing the requirements for data size and data distribution, and effectively reducing the error of independent test sets. It is considered to be one of the most commonly used and best classifiers.

Use LIBSVM in the experiment (please refer to the literature "Chih-Chung Chang, Chih-Jen Lin. LIBSVM: Library for support vector machines [J]. ACM Transactions on IntelligentSystems and Technology (TIST), 2011, 2(3): 1- 27") to build SVMs, LIBSVM is a fast and efficient package for classification and regression, using a one-vs-one strategy to solve multi-classification problems. LIBSVM can not only predict classification results but also provide classification probability information for each test sample. For a k-classification problem, the purpose is to estimate the sample Probability of belonging to each class:

For the one-to-one strategy, _pi is obtained by solving the following optimization problem as follows:

where r _ij is a pairwise probability defined as:

In the experiments of the present invention, SVMs are trained to predict classification results with probabilities, and the probability values of these classification results will be applied to error correction for confusing gestures to determine whether the classification results are directly applied or passed through the present invention. A strategy employed for reclassification.

CNN classifier: Convolutional neural network is a deep feedforward neural network, which uses images directly as the input of the network, does not require manual definition and feature selection, and avoids the link of feature selection and feature extraction in traditional recognition algorithms. At the same time, it also has good fault tolerance, parallel processing ability and self-learning ability.

Instead of using MPCNN, (see the literature "Chih-Chung Chang, Chih-Jen Lin.LIBSVM: Library for support vector machines [J]. ACM Transactions on IntelligentSystems and Technology (TIST), 2011,2(3): 1- 27") The present invention adopts a method described in the document "A.Krizhevsky, S.Ilya, and G.E.Hinton. Imagent classification with deep convolutional neural networks [C]//Advances in Neural Information Processing Systems 2 (NIPS), 2012:1106-1114 The more complex CNN mentioned in " was trained, and the network structure is shown in Figure 4. This network has a total of 8 layers, including 5 convolutional layers and 3 fully connected layers, and the last fully connected layer outputs a 9-dimensional softmax to express predictions for 9 categories. The first convolutional layer performs convolution operation on the input image of 224×224×3 and 96 convolution kernels of 11×11×3 with a step size of 4. The second convolutional layer performs convolution operations on the output of the first layer after response normalization and pooling with 256 5×5×48 convolution kernels. The third convolutional layer uses 384 convolution kernels of 3×3×256 to perform convolution operations with the output of the second layer after normalized kernel pooling. The number of convolution kernels in the fourth convolutional layer is 384 with a size of 3×3×192, and the fifth convolutional layer has 256 convolution kernels with a size of 3×3×192. Each fully connected layer has 4096 neurons. Due to the complex network structure, the present invention adopts the method of enlarging the data set to deal with over-fitting. This method is implemented by randomly sampling 224×224 blocks from a 256×256 image and horizontally mirroring them, and training the neural network on these harvested blocks. Without this approach, the network would suffer severe overfitting, forcing a smaller network, resulting in the inability to use deep features in SVM training.

CNN-SVM hybrid model: The hybrid CNN-SVM model adopted in the present invention replaces the final output layer in CNN with SVM. First, the processed image is passed into the input layer, and the original CNN is trained many times until the training process Converge or reach the maximum number of iterations. Then input the training samples into the trained CNN model to obtain the feature vectors of the training samples, which are input into the SVM classifier for secondary training. After the training is completed, the CNN-SVM model is obtained, and the test samples are input into the model to obtain the classification results.

Error Correction Strategy Based on Probability Estimation (HECS): LIBSVM gives a probability estimate of each sample being classified into each category in the final prediction result, and the final selected classification result is the one with the largest probability value. Table 1 column The final probability distribution of some test samples with wrong predicted classification results is shown. The first column in the table represents the real category number of the test sample, the second column in the table represents the predicted classification number of the test sample, and the remaining columns are respectively Represents the probability that the sample belongs to a certain column, from which it can be observed that in the probability estimation of the wrongly predicted test sample, the maximum value of the estimated probability is the predicted value and the second maximum value is its true value.

Table 1

According to the final decision-making characteristics of LIBSVM and the final experimental results, it can be known that in the samples with wrong predicted classification results, the probability estimation gap between the predicted classification and the real classification is very small, while in the samples with correct predicted results, the predicted classification The gap between the probability estimates and other classification results is relatively large. According to this feature, the present invention proposes an error correction strategy based on probability estimation to reduce classification errors in this case. In the N classification problem, the present invention uses _Mi as a threshold for error correction of all test samples whose prediction result is _i , and the description of Mi is as follows:

Where S _i represents the number of all test samples predicted to be class i, P _n (i) represents the maximum probability estimate of the nth test sample in all pictures predicted to be class i, and P _n (j) represents the second largest value. i represents the class to which the largest value in the probability estimate belongs, and j represents the class to which the second largest value in the probability estimate belongs.

When the probability estimate satisfies the following conditions, modify the class corresponding to the largest value to the class corresponding to the second largest value.

Among them, w _n (i) represents the distance between the predicted maximum value and the second maximum value of the probability estimate of class i, that is, it is equal to P _n (i)-P _n (j) in value, and p _ij represents in the confusion matrix, the prediction Probability of outcome i but true value j.

The advantages of the model of the present invention are as follows:

The invention builds a CNN-SVM model in order to make up for the limitations of CNN and SVM classifiers and combine the advantages of the two classifiers. Theoretical learning method of convolutional neural network and multilayer perceptron (MLP) (refer to "E.A.Zanaty.Support Vector Machines (SVMs) versus Multilayer Perception (MLP) indata classification [J].Egyptian Informatics Journal, 2012,13( 3):177-183"), the learning method is the same, so it is essentially an extension of MLP. The MLP theory is based on empirical risk minimization, which continuously minimizes the training error during the training process. When a minimum value is found during backpropagation calculation, no matter whether it is a global minimum value or not, the training result will converge at this point, and the solution of the algorithm will not continue to be improved. SVM uses the principle of minimizing structural risk to find an optimal hyperplane and minimizes the generalization error on the data when the distribution of the training sample set is fixed. Therefore, the generalization ability of SVM is better than that of MLP.

The advantage of CNN is that it can automatically extract deep-level features of the input image, and the features are still invariant when the input image is moved and distorted to a certain extent. However, manual extraction of features requires careful design. Traditional manual extraction of features in gesture recognition methods (such as the document "Jiang Y.An HMM based approach for video action recognition using motion trajectories[C]//IEEE InternationalConference on Intelligent Control and Information Porcessing, 2010:359-464.”, “Liu Jie, Huang Jin, Han Dongqi, Tian Feng, el at. Template Matching Algorithm for 3D Gesture Recognition[J]. Journal of Computer-Aided Design&Computer Graphics, 2016, 28 (8): 1365-1372 "provided method) ignores the local visual features of the hand, only notices the outline and color information of the gesture, such as the bending of the fingers, the distance between the fingers, which are very important in gesture recognition feature. Artificially designed feature extraction is easy to ignore and lose some features. Therefore, using CNN to extract features can collect more representative and relevant information than traditional methods.

The error correction strategy actually specifies a threshold to screen out the predicted classification results that may be wrong, and then corrects the final classification decision with a certain probability based on the statistical data obtained from the experiment. Classification of samples through the CNN-SVM model has achieved good results, but for some cases where two samples are difficult to distinguish due to occlusion or image quality problems, accurate judgments cannot be made. The corrective method proposed by the present invention The wrong strategy can correct the classification results of confusing samples in the final decision to improve the final overall accuracy.

Carry out experiment and analysis to the inventive method as follows:

Experimental environment: In this experiment, the gesture recognition model runs on the Windows operating system, and the hardware configuration is: Intel(R) Core(TM) i5-6500 processor, NVIDIA GeForceGT730, 8G of memory, and 2G of video memory. CNN network is built by Caffe, and the present invention adopts radial kernel (Gaussian RBF), utilizes LIBSVM software package to realize SVM classifier. All algorithms in the experiment run on Matlab2014a platform.

The experimental results and analysis are as follows:

In the experiment of the present invention, firstly, the color image and the depth image were segmented and preprocessed, and a total of 29,700 gesture images were obtained as the data set of the present invention, of which 27,000 images were used for model training, and 2,700 images for testing. In the CNN training process, the maximum number of iterations is 30,000. It can be seen from Figure 5 that the system has reached convergence when the iteration is about 10,000 times. Finally, the model with 30,000 iterations is used for testing, and the accuracy rate on the test set is 88.35%. %. Then build the CNN-SVM model, replace the last fully connected layer with an SVM classifier, and put the 4096-dimensional feature vector into the SVM for training and testing. In the experiment of the present invention, the SVM adopts the RBF kernel function, in order to find the optimal multiplication coefficient C and the optimal kernel parameter g, the 5-fold cross-validation method is adopted on the training set to obtain the optimal result. The search ranges for these two parameters are: g=[2 ³ , 2 ¹ , . . . , 2 ⁻¹⁵ ] and C=[2 ¹⁵ , 2 ¹³ , . . . , 2 ⁻⁵ ]. A total of 11×10=110 different combinations were tried, and finally C=64, g=0.00024414 were determined. Then the two obtained parameters are used for the training of the hybrid model, and the final training accuracy rate is 99.94%, and the accuracy rate on 2700 test pictures reaches 95.81%. Table 2 lists the training accuracy and test accuracy using CNN and CNN-SVM on the data set prepared by the present invention.

It can be seen from Figure 5 that when the maximum number of iterations is the same as 30,000, the accuracy of the color image is the lowest, at most it can only reach 37.92% accuracy, and the depth image has a significant improvement compared with the color image and can reach 79.07%. , the highest accuracy of the preprocessed image can reach 88.35%. This is because when the unprocessed color image is directly used for training, the training sample itself has a lot of noise information (complex background information and information of other parts of the human body), although the segmented depth image has no background and other parts of the human body. interference, but because the collected depth image is saved by projecting the depth information into the grayscale information of [0,255], so the depth image will also have some information missing, and the gestures after the preprocessing and segmentation of the present invention are not only It can effectively remove complex background and large interference from other parts of the human body while retaining the complete color information of the gesture area, so that more abundant features can be extracted for classification during CNN network training. By putting the test samples into the mixed model for classification prediction, a confusion matrix can be calculated as shown in Table 2:

Table 2

In 100 trials, the error correction rate is mainly concentrated in [3%, 5%], the accuracy rate is most concentrated in [97%, 98%], the average error correction rate is 4.12%, and the average accuracy rate is 97.32%.

Table 3 shows the accuracy rate of gesture recognition in the provided data set by the method of the present invention and other methods. Different from the method of the present invention, the document "Yamashita T, Watasue T. Hand posture recognition based on bottom-up structured deep conbolutional neural network with curriculum learning [C] // Image Processing (ICIP), 2014 IEEE International Conference on. IEEE, 2014: 853- 857" uses a relatively simple convolutional neural network. The maximum pooling layer and convolutional neural network are used to form MPCNN, and the recognition accuracy rate is 68.89% on the test set. Literature "Shao-Zi Li, Bin Yu, WeiWu, Song-Zhi Su, Rong-Rong Ji. Feature learning cased on SAE-PCA network for human gesture recognition in RGBD images [J]. Neurocomputing, 2015, 151(2): 565 -573” uses an end-to-end convolutional neural network, and the gesture recognition accuracy rate is 85.43%. The literature "Xiao-Xiao Niu, Ching Y.Suen.A novel hybrid CNN-SVM classifier for recognizing handwritten digits[J].Pattern Recognition,2012,45(4):1318-1325" first uses depth information and skin color information for gesture segmentation , and then use the SAE-PCA model based on feature learning to extract features, and finally use the SVM classifier to classify. The final accuracy of gesture recognition is 93.32%. The accuracy of different gesture recognition methods on the data set of the present invention is shown in the table 3 shows:

table 3

It can be seen that the method of the present invention has obvious improvement in recognition accuracy compared with other methods.

The method of the present invention first performs segmentation preprocessing on the depth data and color data of the gesture, eliminating the influence of the human body and the complex background of the color data; The complex process of designing features; then estimate the probability of gestures through the support vector machine; finally, based on the obtained probability estimation and the confusion matrix obtained from the experiment, an error correction strategy is proposed to correct the classification results of the model. A large number of experimental results show that this method can effectively recognize static gestures, and can optimize the ability of CNN-SVM model to classify confusing gestures to a certain extent, and can improve the accuracy of final recognition as a whole.

The above-mentioned technical solution is only an embodiment of the present invention. For those skilled in the art, on the basis of the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or deformations, and is not limited to The methods described in the above specific embodiments of the present invention, therefore, the above-described methods are only preferred and not limiting.

Claims (8)

1. a kind of CNN-SVM mixed model gesture identification method based on error correction strategies, it is characterised in that: the method is right first Collected gesture data is pre-processed, and is then automatically extracted feature and is carried out predicting that classification obtains classification results, last benefit The classification results are corrected with error correction strategies.

2. the CNN-SVM mixed model gesture identification method according to claim 1 based on error correction strategies, feature exist In: the described method includes:

Step 1: being pre-processed to obtain training sample and test sample to collected data；

Step 2: obtaining CNN-SVM mixed model；

It is trained step 3: test sample is input in the CNN-SVM mixed model that second step obtains, obtains classification results And the probability Estimation and confusion matrix of classification results；

Step 4: the probability Estimation and confusion matrix that obtain based on third step obtain error correction strategies, error correction strategies are then utilized Classification results are corrected.

3. the CNN-SVM mixed model gesture identification method according to claim 2 based on error correction strategies, feature exist In: the operation of the first step includes:

(11) static gesture is acquired, obtains the depth image and color image of hand respectively；

(12) processing is carried out to the depth image and obtains mask images；

(13) color image and mask images are carried out obtaining coarse gesture area image with operation；

(14) figure after skin color segmentation is divided is carried out to the coarse gesture area image using Bayes's complexion model Image after segmentation is divided into two parts by picture, and a part is used as training sample, and another part is as test sample.

4. the CNN-SVM mixed model gesture identification method according to claim 3 based on error correction strategies, feature exist In: it is that static gesture is acquired using Kinect in the step (11).

5. the CNN-SVM mixed model gesture identification method according to claim 3 based on error correction strategies, feature exist The last output layer that CNN classifier is replaced with SVM classifier is achieved in that in: the second step.

6. the CNN-SVM mixed model gesture identification method according to claim 5 based on error correction strategies, feature exist In: the operation of the second step includes:

(21) training sample is input to the input layer of CNN classifier, by the training of CNN classifier until training process Maximum the number of iterations is restrained or reached, trained CNN model is obtained；

(22): the training sample being input in the trained CNN model and carries out Automatic Feature Extraction acquisition training sample This feature vector；

(23): the feature vector of the training sample being input in SVM classifier and carries out second training, is obtained after the completion of training CNN-SVM mixed model.

7. the CNN-SVM mixed model gesture identification method according to claim 2 based on error correction strategies, feature exist In: the error correction strategies refer to: one threshold value of regulation, are screened the classification results of mistake according to the threshold value, then foundation The statistical data obtained is tested, final classification results are corrected.

8. the CNN-SVM mixed model gesture identification method according to claim 1 based on error correction strategies, feature exist In: the operation of the 4th step includes:

In N classification problem, if M_iFor be i to classification results all test samples carry out error correction a threshold value, for M_i's It is described as follows:

Wherein, M_i,jExpression prediction result is i, but the mean value that true value is calculated by the sample of j, M_iIt is a j dimensional vector； S_i,jExpression prediction result is i, but true value is the quantity of all samples of j, S_iIndicate all test samples for being predicted as i class Quantity, P_n(i) maximum of the probability Estimation of n-th of test sample in all test samples for being predicted as i class is represented Value, P_n(j) second largest value is represented；The class that belongs to of maximum value in the estimation of i presentation class, second largest value belongs in the estimation of j presentation class Class；

When probability Estimation meets the following conditions, the corresponding class of the maximum value of probability Estimation is revised as corresponding to second largest value Class:

Wherein w_n(i) expression prediction result is the probability Estimation maximum value of i class at a distance from probability Estimation second largest value, i.e., in numerical value It is upper to be equal to P_n(i)-P_n(j), p_ijIndicate the probability that classification results are i in confusion matrix but true value is j.