WO2018187952A1

WO2018187952A1 - Kernel discriminant analysis approximation method based on neural network

Info

Publication number: WO2018187952A1
Application number: PCT/CN2017/080176
Authority: WO
Inventors: 邹霞
Original assignee: 邹霞
Priority date: 2017-04-12
Filing date: 2017-04-12
Publication date: 2018-10-18

Abstract

A kernel discriminant analysis approximation method based on a neural network. The method comprises: establishing a training set image set, storing a training set face bitmap, and reading bitmap data; performing characteristic extraction on a training sample in an original input space to form a training set sample set Y; performing characteristic extraction on the training sample set Y to form a training sample set Z; training an RBF neural network using the training set bitmap data and the training set sample set Z formed after the characteristic extraction; establishing a test set image set, storing a test set face bitmap, and reading bitmap data; inputting the test set bitmap data to the trained RBF neural network to obtain a test set sample point set; and classifying and recognizing test set images using a classifier. Compared with the prior art, the kernel discriminant analysis approximation method based on a neural network, and a facial recognition method based on a neural network have a fast approximation speed.

Description

Kernel discriminant analysis approximation method based on neural network

[0001] The present invention relates to a kernel discriminant analysis approximation method based on a neural network, and belongs to the field of face recognition.

Background technique

[0002] Face recognition is a computer technology that achieves the purpose of identity identification by analyzing human facial visual features. The academic community gives a specific definition of face recognition in both broad and narrow sense. Generalized face recognition includes face detection, face representation, face identification, and face expression analysis.

And a series of related technologies such as physical classification; narrow face recognition is defined as a technology or system that enables identity confirmation, identity comparison, and identity lookup through facial features.

[0003] At present, since the face recognition technology can distinguish individuals by the biological characteristics of the organism (generally referred to as a person), and the accuracy of the organism recognition is improved, the technology has been widely concerned and highly respected, and the field has been It has also become a hot spot in the study of biometric features. In humans, for example, biometrics mainly come from the following aspects: face, retina, iris, palmprint, fingerprint, voice, body shape, habits, etc. Therefore, based on the above, research has focused on identifying faces, retinas, and irises. Computer recognition technology for palm embossing, fingerprints, voice, body shape, keyboard tapping, signature, etc., and has achieved significant results.

[0004] The advantage of face recognition lies in its natural and friendly characteristics. The so-called natural nature means that human beings also identify and confirm the identity of each other by observing and comparing human facial features. For example, speech recognition and body shape recognition also have natural features, while humans or other creatures usually do not pass fingerprints. Features such as the iris distinguish individuals, so the above feature recognition does not have natural features.

[0005] The so-called friendliness means that the identification method does not increase the psychological burden of the authenticated person due to special treatment, and thus it is easier to obtain direct and authentic feature information. Fingerprint or iris recognition needs to use special techniques such as electronic pressure sensor or infrared to collect information. The above special collection technology is easy to be discovered, which greatly increases the possibility of the authenticated person avoiding identity identification and reduces the efficiency of identity authentication. [0006] However, face recognition can directly obtain the face information of the authenticated person through simple image or video technology. This information collection method is not easy to be perceived, and the authenticity and reliability of the information are increased.

technical problem

[0007] Although the face recognition technology has the above advantages, the implementation of the technology is not easy. Mainly limited by the biological characteristics of the face, as follows:

[0008] First, since the structures of the same type of faces have high similarities. This feature can be used for face positioning, but it greatly increases the difficulty of using individual facial features to identify individuals.

[0009] Secondly, due to factors such as age, mood, temperature and illumination conditions, and coverings, the shape of the face is very unstable, and even at different viewing angles, the image features of the face are significantly different, increasing the number of people. The complexity of face recognition technology applications.

[0010] In order for the face recognition technology to better serve the required fields, it is necessary to conduct research and breakthroughs in the above two limitations.

[0011] Face recognition technology based on nuclear space is one of the most widely used techniques in the field of face recognition. A general nuclear subspace face recognition method flow is shown in Figure 1. Since all the training samples are used in the representation of the basis in the nuclear subspace, the projection speed of the test sample slows down as the number of training samples increases. It seriously affects the speed of face recognition, especially in the real system and the online system.

Problem solution

Technical solution

In view of the above deficiencies of the prior art, it is an object of the present invention to provide a kernel discriminant analysis approximation method based on a neural network.

[0013] In order to achieve the above object, the present invention adopts the following technical solutions:

[0014] The present invention provides a neural network based kernel discriminant analysis approximation method, including the following steps: [0015] Step 1: Establish a training set image set, store the training set face bitmap, and read the bitmap Data

[0016] Step 2: performing feature extraction on the training samples in the original input space to form a training set sample set Y

[0017] Step three, feature extraction of the training sample set Y, forming a training sample set Z; [0018] Step 4: training an RB F neural network by using the training set bitmap data and the feature set Z of the training set after feature extraction;

[0019] Step 5: Establish a test set image set, store the test set face bitmap, and read the bitmap data [0020] Step 6. Input the test set bitmap data into the trained RBF neural network to obtain the test. Set sample point

[0021] Step 7. Using the classifier, classify and identify the test set image.

[0022] Preferably, in step 2 above, the feature extraction in the original input space is performed by the KPCA method.

[0023] Preferably, the above step 3 performs LDA feature extraction on the training sample set Y.

Advantageous effects of the invention

Beneficial effect

[0024] Compared with the prior art, the neural network-based kernel discriminant analysis approximation method provided by the present invention provides a neural network-based face recognition method, and an online system with a fast approximation speed and a large number of training samples.吋Systems and fingerprint recognition with high recognition speed, license plate recognition and other fields will have broad application prospects.

Brief description of the drawing

DRAWINGS

1 is a schematic flow chart of a general nuclear subspace face recognition method in the prior art;

2 is a schematic flow chart of a kernel discriminant analysis approximation method based on a neural network according to the present invention.

Embodiments of the invention

The present invention provides a kernel discriminant analysis approximation method based on a neural network. The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0028] This embodiment mainly uses the training set image data itself and its nuclear subspace method for feature extraction. If you train an RBF neural network. This RBF neural network is used in the face recognition process, so that the test set image data can be quickly input to the result of approximating the feature extraction of the nuclear space method after inputting the RBF neural network.

[0029] As shown in FIG. 2, the neural network-based kernel discriminant analysis approximation method provided by the present invention specifically includes the following steps:

[0030] (1) Establishing a training set image set, storing the training set face bitmap, and reading the bitmap data.

[0031] (2) Using the KPCA method to perform feature extraction on the training samples in the original input space to form a training set sample set ¥.

[0032] (3) LDA feature extraction is performed on the training sample set Y to form a training sample set Z.

[0033] (4) Training an RBF neural network using the training set bitmap data and the feature set Z of the training set after feature extraction.

[0034] (5) Establishing a test set image set, storing the test set face bitmap, and reading the bitmap data.

[0035] (6) The test set bitmap data is input into the trained RBF neural network to obtain a test set sample point set.

[0036] (7) classifying and identifying the test set image by using a classifier.

Compared with the prior art, the neural network-based kernel discriminant analysis approximation method provided by the present invention provides a neural network-based face recognition method, and the online system with fast approximation speed and large number of training samples is real.吋Systems and fingerprint recognition with high recognition speed, license plate recognition and other fields will have broad application prospects.

[0038]

[0039] It is to be understood that those skilled in the art can make equivalent substitutions or changes in accordance with the technical solutions of the present invention and the inventive concept thereof, and all such changes or substitutions should belong to the appended claims. protected range.

Claims

Claim

[Claim 1] A kernel discriminant analysis approximation method based on a neural network, characterized in that: the approximation method comprises the following steps:

Step 1: Establish a training set image set, store the training set face bitmap, and read the bitmap data;

Step 2: performing feature extraction on the training samples in the original input space to form a training set sample set Y;

Step 3: Perform feature extraction on the training sample set Y to form a training sample set Z; Step 4: Train an RBF neural network by using the training set bitmap data and the feature set sample set after the feature extraction;

Step 5: establishing a test set image set, storing the test set face bitmap, and reading the bitmap data;

Step 6: input the test set bitmap data into the trained RBF neural network to obtain a test set sample point set;

Step 7. Using the classifier, classify and identify the test set image.

[Claim 2] The neural network based kernel discriminant analysis approximation method according to claim 1, wherein: in the second step, the KPCA method performs feature extraction on the training samples in the original input space.

[Claim 3] The neural network based kernel discriminant analysis approximation method according to claim 1, wherein: the step 3 performs LDA feature extraction on the training sample set Y.