CN107480600A - A kind of gesture identification method based on depth convolutional neural networks - Google Patents

Abstract

The invention discloses a kind of gesture identification method based on depth convolutional neural networks, including：(1) division of edge detection process and sample set is carried out to the sample image of training set；（2）Build depth convolutional neural networks；（3）Determine activation primitive and loss function；（4）Train deep neural network；（5）Gesture identification is realized according to the depth convolutional neural networks after training：Its step includes：A) hand-type image is extracted from gesture data to be identified；B) hand-type image is subjected to rim detection and size normalized；C) normalized hand-type image is input in depth convolutional neural networks, the output valve according to output layer judges the ownership class of current gesture.The present invention uses multiple down-sampling technique construction depth convolutional neural networks, and carries out the training of neutral net as activation primitive using hyperbolic tangent function, can not only improve the efficiency of gesture identification, and can improve the accuracy rate of gesture identification.

Description

A Gesture Recognition Method Based on Deep Convolutional Neural Network

technical field

The invention relates to the field of biological feature recognition, in particular to a gesture recognition method based on a deep convolutional neural network.

Background technique

Biometric identification is one of the key technologies in the fields of video surveillance and security authentication. Biological characteristics can be divided into physiological characteristics and behavioral characteristics. Physiological features mainly include face, fingerprints, and irises, while behavioral features include gait, gestures, etc. Typical recognition methods based on physiological characteristics include fingerprint recognition, palm shape and contour recognition, face recognition, iris recognition, etc. Fingerprint recognition is currently one of the most widely used biometric-based identification methods. Fingerprint identification has the advantages of mature technology and low cost. Its disadvantage is that it is contact-type, invasive, has health problems, and fingerprints are also easy to wear. Face recognition technology is a very active research field in recent years. It has the advantages of good intuition, convenience, friendliness, and easy acceptance. Face recognition is non-contact, passive recognition, and does not require the active cooperation of people; but the disadvantage is that it is easily affected by light, viewing angle, occlusion, environment, expression, etc., making recognition difficult. The security and accuracy of iris feature recognition are very high, but the process of feature collection is very difficult. Identification technologies based on behavioral characteristics, common ones are gait recognition and gesture recognition. The input of gait recognition is a sequence of walking video images, which has a large amount of data, resulting in high computational complexity and difficult processing. Gesture recognition is an important part of non-contact human-computer interaction. At present, most researchers focus on the final recognition of gestures, usually simplifying the background of gestures, and using the researched algorithm in a single background. Gestures are segmented, and then the meaning expressed by the gestures is analyzed by the system using commonly used recognition methods.

Contents of the invention

The object of the present invention is to provide a gesture recognition method based on a deep convolutional neural network to address the deficiencies of the prior art.

The object of the present invention is achieved by the following technical solutions: a gesture recognition method based on deep neural network, comprising the following steps:

(1) Carry out edge detection processing and division of sample sets on the sample images of the training set: first, carry out hand type detection and edge detection processing on the gesture set images for training, and adjust the extracted hand type images to a uniform size; The preprocessed data is divided into training sample set and verification sample set;

(2) Constructing a deep convolutional neural network: Let I and O be the input layer and output layer of the deep convolutional neural network respectively, and the hidden layers between I and O are H1, H2, ..., Hn. Wherein, the input layer is the hand image obtained in step (1), the output layer is a gesture feature vector with a length of N, and the hidden layer adopts multiple down-sampling technology, which allows overlapping between sub-blocks that are down-sampled;

(3) Determine the activation function and loss function: select the nonlinear hyperbolic tangent function shown in formula (1) as the activation function of the neuron; select the loss function shown in formula (2). Among them, n is the number of samples in the training set, x is the hand image, y is the gesture feature vector corresponding to x, and θ is the parameter vector;

(4) Training deep neural network: select m training samples from the training sample set, and use the fastest gradient descent method to calculate the gradient; then, use the verification sample set for verification, and when the correct rate exceeds the preset threshold of 99.5%, the training ends, thereby Obtain a deep neural network with a certain weight w and bias item b;

(5) Realize gesture recognition according to the deep convolutional neural network after training: the steps include: a) extracting the hand shape image from the gesture data to be recognized; b) performing edge detection and size normalization processing on the hand shape image; c ) input the hand shape image into the deep convolutional neural network, and determine the attribution of the current gesture according to the output value of the output layer.

The beneficial effects of the present invention are: the gesture recognition method based on the deep convolutional neural network of the present invention uses multiple down-sampling techniques to construct a deep convolutional neural network, and uses the hyperbolic tangent function as the activation function to train the neural network, which not only can Improve the efficiency of gesture recognition, but also improve the accuracy of gesture recognition.

Description of drawings

Fig. 1 is the implementation process of this method;

Figure 2 Schematic diagram of hand image extraction;

Figure 3 Schematic diagram of multiple downsampling;

Fig. 4 hyperbolic tangent function curve;

Figure 5 Error rate data for recognition rate comparison.

detailed description

The present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in FIG. 1 , the gesture recognition method based on the deep convolutional neural network of the present invention is divided into a training stage and a recognition stage.

In the training phase, a deep convolutional neural network is constructed, and the values of its weight and offset parameters are determined using the training set data. Specifically include the following sub-steps:

1. Extract hand shape images from training data: for a frame in gesture interaction, as shown in Figure 2(a), the present invention first extracts rough hand region images according to skin color features, as shown in Figure 2(b) ;Secondly, the image of the hand area is corrected by filtering and fast ecological expansion and erosion algorithm, and a further hand image is obtained, as shown in Figure 2(c);

2. Carry out edge detection and size normalization processing on the hand shape image: carry out edge detection and binarization processing on the hand shape image obtained in step 1 to obtain a preliminary hand shape contour image, as shown in Figure 2 (d); then , pruning and perfecting the contour curve of the hand shape, and adjusting it to a uniform size to obtain the corresponding hand shape image;

3. Constructing a deep convolutional neural network: For the hand image obtained in step 2, the present invention constructs a deep convolutional neural network for gesture recognition. Let I and O be the input layer and output layer of the deep convolutional neural network respectively, and the hidden layers between I and O are H1, H2, ..., Hn. Among them, the input layer is the hand shape image obtained in step 2, the output layer is a gesture feature vector with a length of N, and the hidden layer adopts multiple downsampling technology, which allows overlap between the downsampled blocks, as shown in Figure 3;

4. Determine the activation function and loss function: as shown in Figure 4, select the nonlinear hyperbolic tangent function as the activation function of the neuron; select the square error function shown in formula (2) as the loss function.

5. Train a deep convolutional neural network:

Select m training samples from the training sample set, use the fastest gradient descent method to calculate the gradient, and iteratively optimize the parameters of each hidden layer according to the loss function; then, use the verification sample set for verification, when the correct rate exceeds the preset threshold 99.5% When , end the training, so as to obtain a deep neural network with a certain weight w and bias item b.

In the recognition stage, first extract the hand shape image from the data to be recognized, and perform edge detection and size normalization processing; then, input it into the trained deep convolutional neural network to judge the current gesture category. Finally, the method is compared with the method using data gloves and the sequence similarity detection (SSDA) method, and Fig. 5 shows the experimental data of the error rate.

Claims (1)

1. a kind of gesture identification method based on depth convolutional neural networks, comprises the following steps：

(1) division of edge detection process and sample set is carried out to the sample image of training set：First to the gesture collection of training Image carries out hand-type detection and edge detection process, and by the hand-type Image Adjusting of extraction to unified size；After pre-processing Data be divided into training sample set and checking sample set；

(2) depth convolutional neural networks are built：If I and O are respectively the input layer and output layer of depth convolutional neural networks, I and O Between hidden layer for H1, H2 ..., Hn.Wherein, input layer is the hand-type image that step (1) obtains, and output layer is a length For N gesture feature vector, hidden layer uses multiple down-sampling technology, it is allowed to has between the piecemeal of down-sampling overlapping；

(3) activation primitive and loss function are determined：The non-linear hyperbolic tan shown in formula (1) is selected as neuron Activation primitive；Select the loss function shown in formula (2).Wherein, n is the number of sample in training set, and x is in hand-type image Point, y are the output valve corresponding to x, and θ is parameter vector；

<mrow> <mi>tanh</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mrow> <mn>1</mn> <mo>-</mo> <msup> <mi>e</mi> <mrow> <mo>-</mo> <mn>2</mn> <mi>x</mi> </mrow> </msup> </mrow> <mrow> <mn>1</mn> <mo>+</mo> <msup> <mi>e</mi> <mrow> <mo>-</mo> <mn>2</mn> <mi>x</mi> </mrow> </msup> </mrow> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow>

<mrow> <mi>L</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <munder> <mo>&amp;Sigma;</mo> <mi>x</mi> </munder> <mo>|</mo> <mo>|</mo> <mi>y</mi> <mo>-</mo> <mi>f</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>,</mo> <mi>&amp;theta;</mi> <mo>)</mo> </mrow> <mo>|</mo> <msup> <mo>|</mo> <mn>2</mn> </msup> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2</mn> <mo>)</mo> </mrow> </mrow>

(4) deep neural network is trained：Concentrated from training sample and choose m training sample, calculated using steepest descent method Gradient；Then, verified using checking sample set, when accuracy exceedes predetermined threshold value, terminate training, so as to be had Determine weight w and bias term b deep neural network；

(5) gesture identification is realized according to the depth convolutional neural networks after training：Its step includes：A) from gesture data to be identified Middle extraction hand-type image；B) hand-type image is subjected to rim detection and size normalized；C) hand-type image is input to depth Spend in convolutional neural networks, the output valve according to output layer judges the ownership class of current gesture.