CN107527018A - Momentum Face Detection Method Based on BP Neural Network - Google Patents

Abstract

Gabor characteristic and factor of momentum back-propagation algorithm are combined by the momentum method for detecting human face based on BP neural network, this method.The Gabor characteristic of training set is extracted first, and is entered into factor of momentum reverse transmittance nerve network and is trained.Then, go to detect in input picture using the system trained and whether there is face, if there is then being marked with rectangle.In order to improve the training effect of conventional counter propagation algorithm, factor of momentum is added in the algorithm, effectively slows down concussion trend of the neutral net in training, algorithm can be avoided to be absorbed in local minimum.In addition, increased factor of momentum can be adaptively adjusted the weighted value of every layer of reverse transmittance nerve network.It is substantial amounts of test result indicates that, compared with classics or state-of-the-art Face datection model, our experimental program is effective and has a competitiveness.

Description

Momentum Face Detection Method Based on BP Neural Network

technical field

The invention relates to the field of image data processing, in particular to a momentum face detection method based on BP neural network.

Background technique

Face detection is a worthy research topic in computer vision. In the past many years, researchers have invested a lot of experience in the field of face detection. The essence of face detection is to mark out the faces in the image with rectangular labels. With the increase of face detection applications, face detection has gradually developed into an independent research topic, which has attracted the attention of researchers.

Generally, face detection can be divided into two categories: one is face detection on static images (grayscale or color images), and single or multiple faces can be detected on the image according to the number of faces on the image . The other is face detection on dynamic images, also known as target tracking. The research in this paper is based on face detection in color images. The process of face detection is actually a comprehensive judgment on the characteristics of the face pattern. The input image may contain a large number of pattern features, which can be divided into two categories through color attributes: one is skin color features, and the other is grayscale features.

The application of neural networks to face detection has appeared as early as 1994, followed by the use of convolutional neural networks to train classifiers, retinal connection neural networks to improve frontal face detection, and detection of people with biased angles in images. Face, and the latest convolutional neural network cascading method are used for face detection. In the face detection and recognition of this type of neural network, there are more or less problems with the convergence time of the neural network algorithm being too long, and in the network There is a certain oscillation trend in the training process, which makes the algorithm fall into a local minimum. Eventually lead to the instability of the neural network algorithm.

In the face recognition method, applying the Gabor transform to the neural network has been involved in the prior art, and the Gabor transform is a window Fourier transform. The Gabor function can extract related features in different scales and directions in the frequency domain. The Gabor kernel can extract the features of the image at a specific frequency.

Similar to the Chinese patent number: CN201210057616 based on the local feature Gabor wavelet neural network; the patent of "Face Recognition Method Based on Gabor Wavelet Transform and Local Binary Mode Optimization" applied by Shanghai University, the publication number is CN102024141A. This patent adopts a method based on Gabor wavelet transform and local binary model to optimally fuse together; Tsinghua University applied for the patent of "face recognition method and device for fusion of face part features and Gabor face features", the publication number is CN101276421 , combining the Gabor wavelet transform with the features of face parts. The above-mentioned prior art not only requires a relatively large amount of data calculation, but also takes a long time to converge.

Contents of the invention

The technical problem to be solved by the present invention is: to design a method to solve the problem that the convergence time of the neural network in face detection is relatively long, and it is easy to generate oscillations in training.

In order to solve the problems of the technologies described above, the present invention proposes a kind of momentum face detection method based on BP neural network.

Step 1: Extract the image Gabor features of the training set;

Step 2: Input it into the momentum factor backpropagation neural network for training;

Step 3: Use the trained system to detect whether there is a human face in the input image, and mark it with a rectangle if it exists.

As a preference: the Gabor feature extraction method in step 1 is to select Gabor kernels in five scales and eight directions to extract Gabor features in the image, and convolve the input image with 5*8 Gabor kernels , generating 40 image features of different scales at different frequencies.

Beneficial effects of the present invention:

1. In order to improve the training effect of the traditional backpropagation algorithm, the momentum factor is added to the algorithm, which can effectively slow down the oscillation trend of the neural network during training and prevent the algorithm from falling into a local minimum. In addition, the added momentum factor can adaptively adjust the weight value of each layer of the backpropagation neural network. Extensive experimental results show that our experimental scheme is effective compared with classical or state-of-the-art face detection models.

2. For the Gabor feature extraction of images, 5*8 Gabor kernels are used for convolution, which can effectively extract complex images, especially those with color texture, or images with low color difference and inconspicuous eigenvalues , effectively improving the rapidity of subsequent BP neural network training.

Description of drawings

Accompanying drawing 1: The process of the BP backpropagation algorithm of the method of the present invention.

Accompanying drawing 2: The effect diagram of the detection method of the present invention single face figure.

Accompanying drawing 3: The effect drawing of the method of the present invention to detect multi-face image.

Accompanying drawing 4: The comparison diagram of the average consumption time of the method of the present invention and BP.

Figure 5: Schematic diagram of 40 Gabor filters.

Accompanying drawing 6: utilize 40 Gabor filters in the inventive method to the extraction effect demo figure of the image feature of a pair of portraits;

detailed description

The present invention first selects Gabor kernels on five scales and eight directions to extract the Gabor features in the input image, and convolves the input image with 5*8 Gabor kernels, as shown in Figure 5, using these Gabor kernels Generate 40 image features of different scales at different frequencies, which can also be called 40 Gabor filters. The rows in the figure represent eight different orientations, and the columns represent five different scales. The specific process is as follows:

Given the input image as the corresponding input signal f _in , it is first transformed into the frequency domain by Fourier transform

(x, y) represent coordinates on the spatial domain. The result of the spatial signal is then multiplied by a Fourier transform of a Gabor kernel that can obtain the resulting image filtered by a Gabor filter.

Using the convolution theorem formula (2) is as follows,

Among them, the Gabor kernel is convolved with the input signal to obtain the response of the input signal close to a certain neighborhood.

Figure 6 shows an application of formula (2), where 6-(a) is the input original image. 6-(b) shows 5*8 different frequencies and different scales of Gabor kernels (ie: 40 Gabor filters) to obtain the response of the original image 6-(a) at different frequencies and different scales. The responses are extracted as image features.

The two-dimensional complex wave represented by formula (3) is multiplied by the two-dimensional Gaussian function calculated in formula (4), so as to obtain the Gabor kernel in formula (5).

s(x,y)=exp(i(2π(u ₀ x+v ₀ y))+p) (3)

Among them, the initial phase p has little influence on the Gabor kernel and can be omitted.

δ _x and δ _y respectively control the "spread" of the Gaussian function in the x and y directions.

(x ₀ ,y ₀ ) is the center point of the Gaussian kernel, θ is the rotation direction of the Gaussian kernel, (δ _x ,δ _y ) is the scale of the Gaussian kernel in the x,y direction, (u ₀ ,v ₀ ) is the frequency Domain coordinates, K is the magnitude scale of the Gaussian kernel.

The widely used BP neural network is used as the basis of the present invention. In order to overcome the shortcomings of the BP algorithm, the momentum item can be used to improve the convergence speed of the algorithm and avoid the algorithm from falling into a local minimum. The overall steps of the neural network algorithm in the face detection system we propose are shown in Figure 1, which briefly shows the workflow of the neural network. Next, the momentum factor backpropagation neural network algorithm is introduced in detail.

The extracted image Gabor feature 1 is used as the input of the neural network, which is a fully connected network structure. As shown in Figure 1, one layer is the input layer 2 and the other layer is the hidden layer 3.

The transfer function of the neurons in the hidden layer is shown in the following formula (6):

In this function, net is the input data from the input layer. Each input is regarded as x _i , if given n input data, among them, 1≤i≤n, then the formula of net is as shown in (7):

net＝x ₁ w ₁ +x ₂ w ₂ +...+x _n w _n (7)

Among them, w _i is the initial weight of the neural network. In addition, in the neural network, the number of nodes in the input layer and output layer is known, and the number h of nodes in the hidden layer is determined by formula (8).

Among them, m and n represent the number of nodes in the input layer and output layer respectively, and A represents an adjustable constant value between 1 and 10.

Figure 1 adds a momentum factor as a feedback parameter through the output value 4 of the BP neural network, and increases the feedback parameter when forwarding the feedback error signal to improve the training performance of the traditional BP neural network.

The process of forward propagation can be calculated by formula (9) as follows.

Wherein, x _j =f(S _j ), net=S _j .

The core idea of BP is to pass the output error back to the input layer layer by layer through the hidden layer in some form. The initial weight w can be used during the forward propagation of the backpropagation network, which is generated by random initialization. However, compared with the ideal output, the actual output may have a large error. In order to continuously adjust w, the error function can be obtained in formula (10), where y _j represents the actual output.

The input training set is trained in the following computational steps during training.

1. Calculate the error term formula (11) in each unit of each layer from the reverse output layer.

error _i =y _j (1-y _j )(d _j -y _j ) (11)

2. Calculate the error formula (12) of the hidden layer nodes.

error _h = y _h (1-y _h ) error _h (12)

3. Update each weight formula (13).

w _ik ＝w _ik +μ error _k x _ik (13)

Wherein, Δwi _ik =μ·error _k ·xi _ik is the update rule of the weight, x _ik represents the input value, and wi _ik is the corresponding weight value between node i and node k.

These are the most basic mathematical ideas in neural networks, which can be trained on positive examples and negative examples, with relatively long time consumption. The following is the improved formula (14) for the neural network.

Δw _ik (n)=μδ _k x _ik +αΔw _ik (n-1) (14)

Where 0≤α<1 is the momentum term. And α is a global parameter, which can be determined by the output value of the test and output layer. When the gradient keeps the same direction, the step size can be increased and the iterative path can be directed to the minimum target value. Usually it is necessary to reduce the learning parameter μ when α is large (that is, α is close to 1). If the gradient direction is always changing, the addition of the momentum term can make the change of the iterative path smooth.

This is mainly due to the poor training of the neural network, that is, the curvature in different directions is not the same, which forms long and narrow surface valleys of different sizes. For the vast majority of points on the surface valley, the gradient of the iterative path does not point to the direction of the minimum value, so the iteration path oscillates from one side to the other between the continuous steps of the iterative path, and the speed of convergence to the minimum value is also greatly improved. Slower, if the momentum item is added, the shock can be effectively weakened and the convergence speed can be greatly improved.

In formula (14), the weight of the nth iteration depends on the weight of the n-1th iteration. Increasing the momentum term to a certain extent improves the effect of the search step, which can make the algorithm converge faster. On the other hand, since the multi-layer network may cause the algorithm to converge to a local minimum due to the loss function, the momentum item can cross some local minimums to some extent to avoid the algorithm from falling into the local minimum.

Experimental verification:

The Face Detection Data Set face dataset and the CMU and Harvard face database containing about 5000 frontal faces are selected as the face library of the training set. Among them, the CMU and Harvard face database contains about 5000 face samples, and the faces in these face images have some offset angles and simple backgrounds. At the same time, we also used 250 landscape images without faces to obtain about 2500 non-face images as the non-face library of the training set by bootstrapping.

Figures 2 and 3 show some resulting images after face detection on the Partheenpan test dataset using momentum face detection. In 2-(a) of Fig. 2, we choose frontal face images as input. In 3-(a) of Fig. 3, each input image contains multiple faces. In 2-(b) of Fig. 2 and 3-(b) of Fig. 3, the detected faces are marked with rectangles in the output image.

Our trained momentum face detection system to test some common face databases:

1 Partheenpan Data Set (Partheenpan),

2 The Annotated Faces in the Wild (AFW),

3The Face Detection Data Set&Benchmark(FDD-B)

4CMU Data set (CMU)

In Fig. 4, a comparison is made between the average detection time of the method of the present invention and the average time of detecting human faces in the above four kinds of data by the classical BP neural algorithm.

Because each image in the two data sets of AFW and CMU will have multiple faces and more complex image backgrounds, the performance improvement in the detection of these two data sets is more obvious.

Table 1 below gives the detection rates and false detections on four different face databases. Including the limitations brought by the variety of image sizes and image clarity, the inventive method has a certain false detection rate, but it is in the lead in the prior art.

Table 1 The detection rate and false detection of four groups of different face database test sets

For example, in the AFW test set, the detection rate=(451-65)/451=85.6%, and the false detection rate=39/(39+451)=7.95%.

The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the industry should understand that the present invention is not limited by the above-mentioned embodiments, and that described in the above-mentioned embodiments and the description only illustrates the principles of the present invention, and the present invention also has various aspects without departing from the spirit and scope of the present invention. Variations and improvements all fall within the scope of the claimed invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims (3)

1. the momentum method for detecting human face based on BP neural network, it is characterised in that：

Step 1：Extract the image Gabor characteristic of training set；

Step 2：It is entered into factor of momentum reverse transmittance nerve network and is trained；

Step 3：Go to detect in input picture using the system trained and whether there is face, if there is then being marked with rectangle.

2. according to the method for claim 1, it is characterised in that：In described step 1 Gabor characteristic extraction method be It has selected the Gabor cores on five yardsticks and eight directions to be used for extracting the Gabor characteristic in image, by input picture 5*8 Individual Gabor cores carry out convolution, generate the characteristics of image of 40 different scales under different frequency.

3. according to the method for claim 1, it is characterised in that：Factor of momentum in described step 2 is Δ w_ik(n)=μ δ_kx_ik+αΔw_ik(n-1) wherein 0≤α ＜ 1 are momentum terms；In above formula, the weight of nth iteration depends on (n-1)th iteration Weight.