A method for recognizing image depth

Abstract

The present invention proposes a deep face recognition method, including a method for deep extraction of image information, including a method for large-scale image segmentation of sample image information and a method for large-scale image transformation of sample images. The method also includes a method for deep image recognition, including considering the effect of each element of a large-scale feature set on the recognition result, considering the common probability distribution information of complex feature values, considering the common probability distribution of sample feature values ​​at the same position as the probability distribution of the logged-in feature values, considering the probability information of the maximum probability value between the sample feature vector and the logged-in feature vector, etc. The effect of the present invention is that it can quantitatively evaluate image recognition, make full use of all available information, and thus achieve optimization of face recognition.

Description

[Technical field]

[0001] The invention relates to a method for deep face recognition in the field of image processing.

[0002] At present, the technical competition in the field of face recognition has become the main battlefield of artificial intelligence, and the degree of competition has reached a very fierce level. A representative patent in Japan is the patent published by NEC Corporation entitled: "Pattern Recognition Method" (Patent Document 1). This patent discloses a feature vector composed of 30 eigenvalues ​​of the facial features, which is then used to train the covariance between the set of facial feature vectors of the same person and the set of different facial feature vectors. Finally, through the Mahalanobis distance formula, the facial data corresponding to the feature vector with the minimum distance between the sample feature vector and the different facial feature vectors after login is set as the recognition result.

[0003] NEC Corporation, which has the highest level of facial recognition technology in Japan, later disclosed: "Facial feature data generation method and device, and facial similarity calculation method and device" (Patent Document 2). This patent further adds a reliability index to the recognition of each facial image based on the previous patent. This reliability index is derived from the normal distribution of the vector difference between feature vectors, the variance between the eigenvalues ​​of the difference vector of the feature vector, the post-distribution of the difference vector between facial features as a mixed distribution, derived from the logarithmic degree of the mixed distribution, etc., and the similarity between patterns is calculated according to the reliability, thereby improving the accuracy of facial recognition.

[0004] In 2004, NEC and Samsung Electronics of South Korea also disclosed a patent for face recognition: "Face recognition and feature vector extraction method and device for retrieval" (Patent Document 3). This patent discloses the first normalized vector and the second normalized vector for the entire face image, which are used to generate the overall Fourier transform feature vector through Fourier transform, and then the third normalized vector and the fourth normalized vector for the face image in the center are used to generate the Fourier transform feature vector of the center through Fourier transform. An overall intensity feature vector is generated for the overall face image, and a local intensity feature vector is generated for the face image in a specified local area. The above-mentioned first normalized vector and the second normalized vector are combined with the overall intensity feature vector to generate an overall feature vector, and the above-mentioned third normalized vector and the fourth normalized vector are combined with the local intensity feature vector to generate a local feature vector. This constitutes the feature vector used for retrieval.

[0005] In response to the above invention, NEC Corporation published a patent entitled "Method and device for generating facial feature data, and method and system for facial recognition" in 2006 (Patent Document 4). Based on Patent Document 3, this patent proposes a new calculation method for reflecting the reliability of the original image state, such as generating a reliability value by different image contrasts, generating a reliability value by different maximum grayscale values ​​and minimum grayscale values ​​of pixels in an image, generating a reliability value by different pixel dispersion and standard deviation of an image, generating a reliability value by different asymmetric indicators of an image, and generating a reliability value by different values ​​obtained by multiplying or adding pixel values ​​of a differential image between an extracted image and an inverted image.

[0006] More than a decade later, NEC Corporation published the latest patent on face recognition technology: "Face recognition device, face recognition method, face recognition program, display control device, display control method, and display control program" (Patent Document 5). This patent proposes a method for identifying a face image based on its local part, and derives three results between each face image based on the threshold of the similarity of the face image: similarity, ambiguity, and dissimilarity. These three conclusions of each face image are used as the basis for the recognition of the overall face image.

[0007] [Public Documents]

[0008] [Patent Document 1] Japanese Patent Application No. 9-134432

[0009] [Patent Document 2] Japanese Patent Application No. 2003-187229

[0010] [Patent document 3] Tekai Ping Tekai 2004-139596 (P2004-139596A)

[0011] [Patent Document 4] Japanese Patent Application Laid-Open No. 2006-344236 (P2006-344236A)

[0012] [Patent Document 5] Japanese Patent Application Laid-Open No. 2018-49655 (P2018-49655A)

[0013] The method described in the above-mentioned patent document 1 is used to train the covariance within the set of facial feature vectors of the same person and between different sets of facial feature vectors, and use the Mahalanobis distance to obtain the final recognition result. Although the probability statistics formula algorithm can be applied to face recognition, which can be a great improvement over the traditional image recognition method, there is a large amount of random interference in the face recognition image. How to go beyond the probability statistics formula algorithm, filter out the noise of the face image and the feature value, and make the predicted recognition result close to the matrix is ​​still a long way off. In addition, the Mahalanobis distance can only consider the similarity relationship between feature vectors, and cannot directly use the distance that can unify the Euclidean space and the probability space to calculate the scale relationship between different feature vectors with high precision, which becomes a problem that affects the improvement of recognition accuracy.

[0014] The method described in the above-mentioned patent document 2 adds a reliability index processing for the recognition of each face image. This method can undoubtedly improve the accuracy of face recognition, but the definition of reliability lacks strict scientificity and cannot achieve optimal recognition results.

[0015] The method described in the above-mentioned patent document 3 proposes to use the spectrum of Fourier transform as the feature vector of the face image. Since the rotation of the face image has a great influence on the spectrum, and the spectrum information is not very sensitive to the feature information of the face image, and the feature values ​​are filled with a lot of noise, it seriously affects the improvement of the recognition rate.

[0016] The method described in the above-mentioned Patent Document 4 increases the accuracy of face recognition by generating a reliability that can reflect the state of the original image. However, the generation of the reliability is not rigorous enough.

[0017] The method described in the above-mentioned patent document 5 is a method of identifying the parts of the face image one by one. The final reliability processing uses three reliability thresholds as the judgment basis, which lacks rigor and still has the above-mentioned problems in the recognition of partial face images. [Summary of the invention]

[0018] The first purpose of the present invention is to provide an information extraction method for image recognition that is suitable for generating large-scale feature vectors of the order of billions, which can solve the problem of misalignment of the recognized image and can most effectively extract the feature information of the recognized image.

[0019] The second purpose of the present invention is to provide a method for solving the problem of how to obtain high-confidence feature value data from a large number of feature values ​​of identified images in large-scale image feature vectors.

[0020] The third purpose of the present invention is to propose a method for constructing a new generation of artificial intelligence models with big data feature vectors that can produce a large number of data sets that exceeds the number of traditional deep learning data sets through small data training.

[0021] The fourth objective of the present invention is to propose a model for multi-purpose automatic machine learning of image processing and image recognition.

[0022] The fifth objective of the present invention is to propose a method for 3D object recognition based on automatic machine learning without the need for training.

[0023] The sixth objective of the present invention is to propose a processing method for realizing deep data mining and machine understanding of radar data.

[0024] The present invention adopts the following technical solutions to solve the technical problems:

[0025] A method for extracting depth from image information, characterized in that:

[0026] A method for performing large-scale image segmentation on sample image information, or a method for describing a specific position of an image in detail, or a method for deeply extracting image information by image transformation for at least one segmented image.

[0027] Moreover, the image transformation method transforms the sample image through an orthogonal transformation method including Fourier transform (DFT, Discrete Fourier Transform), discrete cosine transform (DCT, Discrete Cosine Transform), Walsh transform (WT, Walsh, Transform), Hadamard transform (ST, Hadamard Transform), KL transform (KL, Karhunen-Loeve Transform) cotransform (ST, Slant Transform), discrete wavelet transform (DWT, Discrete Wavelet Transform); or an image transformation method including principal component analysis (PCA, Principal Component Analysis), linear discriminant analysis (LDA, Linear Discriminant Analysis), and Gabor transform (GT, Gabor, Transform).

[0028] The present invention also proposes an image depth recognition method, which is characterized by:

[0029] A method for deep image recognition that considers the effect of a large-scale logged-in feature set on recognition results, or a method that considers the ratio between the minimum distance between feature vectors and the cumulative value of the scale of the maximum probability space of each eigenvalue of the feature vector, or a method that considers the common probability distribution information of complex eigenvalues, or a method that considers the probability distribution information of sample eigenvalues ​​at the position of the probability distribution of logged-in eigenvalues, or a method that considers the common probability distribution information of sample eigenvalues ​​at the common position of the probability distribution of logged-in eigenvalues, or a method that considers the probability information of the maximum probability between the sample feature vector and the logged-in feature vector, or a method that considers segmenting the probability distribution of the eigenvalues ​​of the logged-in feature vector.

[0030] Moreover, the method of performing large-scale image transformation on the image also refers to: using multiple image processing methods to perform multiple image processing on at least one sample image; obtaining multiple expanded images produced by different environments that can simulate the sample image; using the multiple expanded images as a set of sample data for the image; and a method for making an image also generate a probability distribution of sample feature values.

[0031] Moreover, the above-mentioned image depth recognition method also includes one of the forward recognition, reverse recognition, or bidirectional recognition methods for the image.

[0032] The advantages and positive effects of the present invention are:

[0033] It is possible to quantitatively evaluate face recognition, extract the depth information that can be obtained through the image, and utilize all the information with deep image recognition, thereby achieving quantitative evaluation of image recognition accuracy and optimizing the recognition results.

[0034] A plurality of simulated multi-scene photographing images can be directly generated for the face photo on the ID card, thereby improving the image recognition accuracy.

[0035] The special terms appearing in this manual are defined as follows:

[0036] The so-called maximum probability value: the statistical expectation value that goes beyond the traditional statistical formula algorithm can be as close as possible to the result of the machine learning processing of the parent.

[0037] The so-called probability scale self-organization: self-organizing iterative processing with the maximum probability space as the scale can obtain the maximum probability value, the maximum probability space and the scale of the maximum probability space.

[0038] The so-called multi-purpose automatic machine learning is an automatic machine learning model that can obtain the optimal solution of the objective function for multiple purposes without training.

[0039] The so-called SDL (Super Deep Learning) model is an artificial intelligence model composed of a new type of neural network dominated by function probability distribution and using complex machine learning to connect nodes.

[0040] The so-called EPD distance: can unify the distance scales of Euclidean space and probability space.

[0041] The so-called FPM measure: a fuzzy event probability measure that can unify Euclidean space and probability space.

Brief Description of the Drawings

[0042] Figure 1 is the given image of the object to be identified

[0043] Figure 2 This is a schematic diagram of the amplitude spectrum image.

[0044] Figure 3 This is a diagram of a centralized data processing method.

[0045] Figure 4 Yes Figure 2 Diagram of data swapping diagonal data

[0046] Figure 5 This is a schematic diagram of the transformation of the two-dimensional phase spectrum.

[0047] Figure 6 This is a schematic diagram of the SDL model that can handle large-scale feature vectors.

[0048] Figure 7 Gray value adjustment image processing method

[0049] Figure 8 It is an image processing method that emphasizes border information

[0050] Fig. 9 It is another image processing method that emphasizes border information.

[0051] Fig.10 A schematic diagram of a single photo image versus a single recognition image

[0052] Fig.11 It is a schematic diagram of the recognition of a single-shot image with simulated multi-scene photo images

[0053] Fig.12 This is a schematic diagram of image recognition of a single photo in multiple scenes.

[0054] Fig.13 Schematic diagram of single photo image recognition for multi-scene shooting and simulated multi-scene shooting

[0055] Fig.14 It is a schematic diagram of dividing a probability space into multiple probability spaces.

[0056] Fig.15 is a diagram of two probability spaces in Euclidean space

[0057] Fig.16 A schematic diagram of a pattern recognition method that uses common probability distribution information

[0058] Fig.17 This is a diagram of a new probability value definition method that takes into account the overall probability.

[0059] Fig.18 This is a process flow chart for multi-purpose automatic machine learning.

[0060] Fig.19 It is a flow chart of the method for machine understanding radar data processing [Specific implementation method]

[0061] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings, but the embodiments of the present invention are illustrative rather than restrictive.

[0062] The present invention solves its technical problem by adopting the following technical solutions, and the embodiments of the invention are described according to the drawings in the specification.

[0063] Firstly, a method for extracting depth information from images is proposed.

[0064] Figure 1 is the given image of the object to be identified.

[0065] against Figure 1 The image can be used to perform discrete Fourier transform (DFT) of a two-dimensional image using the following formula.

[0066] [Formula 1]

[0067]

[0068] Among them, f(x, y) is a digital image of size M*N, u and v are used to determine their frequencies, and the frequency domain system is the coordinate system generated by it.

[0069] The Inverse Discrete Fourier Transform (IDFT) is given by:

[0070] [Formula 2]

[0071]

[0072] right Figure 1 The data after image transformation (original):

[0073]

Data 1

[0074] F(0,0),F(0,1),...F(0,N)

[0075] F(1,0), F(1,1),...F(1,N) ... ...

[0078] F(M,0),F(0,0),...F(M,N)

[0079] Amplitude spectrum calculation method (Magnitude)

[0080] Let R and I represent the real and imaginary parts of F respectively, then the amplitude spectrum formula is as follows:

[0081] [Formula 3]

[0082]

[0083] Since the frequency domain values ​​differ greatly and cannot be displayed directly, they are converted using the following formula 4, where white represents high values.

[0084] [Formula 4]

[0085] M 1 =log(1+M)

[0086] Figure 2 A schematic diagram of the amplitude spectrum image.

[0087] like Figure 2 shown; Figure 1 Applying formula 3 and formula 4 to image data, we can get Figure 2 The amplitude spectrum image shown, here, white represents high value experimental results.

[0088] Figure 3 This is a schematic diagram of the centralized data processing method.

[0089] like Figure 3 shown; Figure 2 By swapping the diagonal data 1 and 3, and 2 and 4, the DC component can be moved to the center of the spectrum. Figure 4 of the image.

[0090] Figure 5 It is a schematic diagram of the transformation of the two-dimensional phase spectrum (Phase).

[0091] like Figure 5 As shown; let R and I represent the real part and imaginary part of F respectively, then the phase spectrum formula is as follows:

[0092] Formula 5

[0093]

[0094] The specific meaning of the face value in the phase spectrum is radians, and the range is: -3.14159 to 3.14159.

[0095] It is a well-known technology to use Fourier transform spectrum for face recognition. Generally, the Fourier transform spectrum will not be affected when the face image is translated. However, when the face image is rotated, it exceeds the discrete window of Fourier transform, so the result of Fourier transform will be wrong. Therefore, it is necessary to process the face image so that it can be included in the discrete window even after rotation.

[0096] The above-mentioned eigenvalue extraction method is one of the methods of image transformation. In addition, as a method of considering image orthogonal transformation, you can also consider discrete cosine transform (DCT, Discrete Cosine Transform), Walsh transform (WT, Walsh, Transform), Hadamard transform (ST, Hadamard Transform), KL transform (KL, Karhunen-Loeve Transform) cotransform (ST, Slant Transform), discrete wavelet transform (DWT, Discrete Wavelet Transform) and so on.

[0097] In addition, you can also consider principal component analysis (PCA, Principal Component Analysis), linear discriminant analysis (LDA, Linear Discriminant Analysis), periodic parameter estimation method (ARS proposed by Yukihiro Kamiya), and Gabor transform (GT, Gabor, Transform) and other methods.

[0098] Here, a method of extracting depth information at a specific position in an image is proposed.

[0099] The face image is transformed into various images through the physical model. For example, the power model is substituted into the face image to generate the power image, and then the information of the transformed face image is deeply extracted. Here, not only the position information of the facial features is extracted, but also each facial feature position should be described more delicately, such as the length information, width information, color information, wrinkle information, the ratio of the upper and lower lips, etc. of the mouth, the size information of the eye, the size information of the eyeball, the color information, the position information of the nose, the shape of the nose, the size of the nose, the color distribution of the nose, etc., the distribution of facial wrinkles, the skin color of the face, the local roughness, the shape of the face, etc. The geometric features of each position of the face, the physical features are all used as feature values ​​for deep extraction of image information.

[0100] Figure 6 is a schematic diagram of the SDL model that can handle large-scale feature vectors.

[0101] like Figure 6 As shown in the figure, this is a new three-layer neural network consisting of a perception layer, a neural layer, and a cerebral cortex. It is an SDL model that can process large-scale feature vectors.

[0102] Here, a method for performing large-scale image transformation is proposed. The image can be segmented on a large scale, and then each image segmentation is transformed, so that millions or even hundreds of millions of feature sets can be obtained to achieve high-precision image recognition results that surpass traditional deep learning.

[0103] like Figure 6 As shown; (FP 1 ) is a part of a face image, such as an image of a person’s eyes. (FP 2 ) is the image of other parts of the face image, for example, the image of the nose. The face image can also be divided into n regions, (FP n ) is the last part of the face image, for example, the image of the mouth. The more detailed the face image segmentation is, the greater the accuracy of face recognition will be.

[0104] When segmenting a face image, the position information of the face image can also be used as the face positioning information for image segmentation, which can solve the problem that the accuracy of the recognition result is affected by the displacement and rotation of the image.

[0105] (IT 1 ), (IT 2 )~(IT m) is a type of image transform, including Fourier transform (DFT, Discrete Fourier Transform), discrete cosine transform (DCT, Discrete Cosine Transform), Walsh transform (WT, Walsh, Transform), Hadamard transform (ST, Hadamard Transform), KL transform (KL, Karhunen-Loeve Transform) cotransform (ST, Slant Transform), discrete wavelet transform (DWT, Discrete Wavelet Transform), principal component analysis (PCA, Principal Component Analysis), linear discriminant analysis (LDA, Linear Discriminant Analysis), periodic parameter estimation method (ARS proposed by Yukihiro Kamiya), and Gabor (GT, Gabor, Transform) transform and other methods.

[0106] like Figure 6 As shown; (ML 1 ') is a probability scale self-organizing machine learning unit, which is responsible for obtaining the feature value with the maximum probability from the multiple feature information and inputting it into the node of the perception layer (N 1 '), when the number of eigenvalues ​​is odd (ML 1 ') can be omitted, and the feature values ​​can be directly input into the nodes of the perception layer (N 1 ')superior.

[0107] In this SDL model that can process large-scale feature vectors, the perception layer (P 1 ') and neural layer (P 2 ') of nodes (N 1 ') and (N 2 ') can be increased to infinity (N 1 (1) ')~(N I (n) '), (N 2 (1) ')~(N 2 (n) '), where n→∞. Therefore, as the number of nodes increases, the probability scale self-organizing machine learning unit also becomes (ML 1 (1) ')~(ML 1 (n) '), (ML 2 (1) ')~(ML 2 (n)'), similarly n→∞.

[0108] After studying the mechanism of deep learning, it is found that the application effect of deep learning depends on the size of huge data sets. However, the cost of increasing the data set for deep learning is O(n n ), this computational complexity limits the indefinite expansion of the dataset size for deep learning. In contrast, the SDL model has a computational complexity of O(n 2 ) unchanged, so it can always be expanded by expanding the perception layer (P 1 ') and neural layer (P 2 ') on the node (N 1 ') and (N 2 '), realize the small data training characteristics of the probability distribution of the function, and the characteristics of obtaining large-scale feature vectors, to achieve application effects beyond deep learning.

[0109] Different from the publicly available SDL model, the image to be recognized by the traditional SDL model is expanded into a mapping image through a physical model. This method makes it difficult to make the feature vector reach the scale of millions or even billions. 1 )~(IT m ) combined with the SDL model, the size of the feature vector can be easily increased greatly. Its purpose is to surpass the size of the deep learning data set, thereby producing application effects beyond deep learning.

[0110] In practical applications, in the face image data, each person often only has a photo image obtained from the ID card. How to use this photo image to generate a probability distribution with the maximum probability value and the maximum probability scale database is the key to importing the SDL model to improve face recognition.

[0111] The present invention proposes to use various image processing methods that can improve the image recognition effect, transform a face image into a plurality of expanded images of face images produced by different environments that can simulate the sample image, and use the plurality of expanded images as a set of sample data of the image; a method for making an image also generate a probability distribution of sample feature values.

[0112] The specific image processing method proposed here is only an example. It is within the scope of the present invention to transform an image into multiple images by using various image processing methods that can improve the image recognition effect and generate a probability distribution of a sample feature vector for image recognition.

[0113] The following is an image processing method for adjusting grayscale values.

[0114] like Figure 7 (a) is the original grayscale value of any 3*3 pixel in the original face image. Figure 7 (b) Exchange the maximum grayscale value and the center grayscale value in the original grayscale value of 3*3 pixels. Figure 7 (c) Exchange the minimum grayscale value of the original grayscale value of 3*3 pixels with the center grayscale value. Figure 7 (d) The gray values ​​of the original gray values ​​of 3*3 pixels are replaced by the central gray value by replacing the high probability value with the probability scale. Figure 7 (e) Exchange the maximum grayscale value and the center grayscale value in the original grayscale values ​​of 3*3 pixels.

[0115] Following the above method, you can also Figure 7 (a) The diagonal line and the center cross line of the original grayscale value of 3*3 pixels are replaced by the maximum grayscale value and the minimum grayscale value in the original grayscale value of 3*3 pixels, and the maximum probability value is obtained by self-organization of the probability scale.

[0116] Figure 8 It is an image processing method that emphasizes border information.

[0117] like Figure 8 As shown in (a); the image is derived in the x and y directions respectively, and then Figure 8 The constants in the left and right 3*3 grids of (a) are multiplied according to the correspondence of each pixel, and the result replaces the grayscale value of the original pixel. Figure 8 (b) of the image; the image is differentiated in the x and y directions respectively, and then Figure 8 (b) The grayscale value of the original pixel is replaced by the result of multiplying the constants in the left and right 3*3 grids.

[0118] Fig. 9 It is another image processing method that emphasizes border information.

[0119] same Figure 8 Same, such as Fig. 9 As shown in (a), by multiplying the x-direction derivative result with this template, we can get the processing effect of the horizontal border filter. Fig. 9 As shown in (b), by multiplying the derivative result in the y direction with this template, the processing effect of the horizontal border filter can be obtained.

[0120] In addition to transforming a face image into multiple images, it is also important to emphasize that the quality of the image can be improved.

[0121] In image processing, in order to improve the visual quality of the image, image brightness transformation, color transformation, image enhancement, suppression of certain components, geometric transformation, etc. are used.

[0122] In addition, by strengthening the high-frequency components of the image, the facial features in the image can be made clear and the details obvious, and by strengthening the low-frequency components of the image, the noise in the image can be removed.

[0123] The image can also be divided into multiple images through image filtering, which can improve the scale of the feature vector while improving the image quality.

[0124] For example, the grayscale value transformation functions include: logarithmic transformation, gamma transformation, piecewise linear transformation, histogram transformation, etc.

[0125] Spatial filtering includes: smoothing spatial filtering, statistical sorting filtering, sharpening spatial filtering, unsharpening concealment and high-lift filtering, gradient sharpening, etc.

[0126] Frequency based filtering includes ideal low pass filtering, ideal high pass filtering, Butel low pass filtering, Butel high pass filtering, Gaussian low pass filtering, Gaussian high pass filtering, passivation template, high boost filtering, high frequency emphasis filtering, homomorphic filtering, selective filtering, and 2-trap filtering.

[0127] The processing of noise probability density function includes: Gaussian noise, Rayleigh noise, Irish noise, exponential noise, uniform noise, bibasic impulse noise and so on.

[0128] Spatial filtering to restore noise includes: mean filter, arithmetic mean filter, geometric mean filter, harmonic mean filter, inverse harmonic mean filter, statistical sorting filter, adaptive filtering, frequency domain elimination of periodic noise, etc.

[0129] In order to improve the accuracy of face recognition, the present invention also proposes the following method related to recognition conditions.

[0130] Fig.10 It is a schematic diagram of a single photo image versus a single recognition image.

[0131] like Fig.10 As shown in the figure, more than 10,000 face images in the database, one face image for each person, are used to extract a feature vector (PV) using the above-mentioned image transformation method or the method of facial features position, or using two methods at the same time, or more methods. max1～n ), log in the feature vector, then use the camera to take a face image of the logged-in photo, and use the above-mentioned image transformation method or the method of facial features position, or use two methods at the same time, or more methods to extract a feature vector (sp_PV max1～n ) to perform face recognition.

[0132] Here, a method for performing positive recognition on images is proposed.

[0133] Fig.11 It is a schematic diagram of the recognition of a single-shot image with simulated multi-scene photo images

[0134] like Fig.11 As shown in the figure, more than 10,000 face images in the database are used, one for each person, and the above image processing method is used to simulate images taken in multiple scenes, so that several images simulating different scenes can be obtained. Then, a group of feature vectors are extracted using the above image transformation method or the method of the position of the facial features, or two or more methods are used at the same time, and the probability scale self-organization processing of the feature vectors is performed to obtain the maximum probability value (PV) generated by the probability scale self-organization. max1～n ) is the expected value of the maximum probability space in statistics, the scale of the maximum probability space (PS max1～n ) is the variance of the maximum probability space, and is logged into the database. Then, a face image of a logged-in photo is taken with a camera, and a feature vector (sp_PV) is extracted using the above-mentioned image transformation method or the method of facial features position, or using two or more methods at the same time. max1～n ) to perform face recognition.

[0135] Alternatively, by directly taking multiple photos of a person, we can obtain images of several different scenes, and then use the above-mentioned image transformation method or the method of facial features position, or use two or more methods at the same time to extract a set of feature vectors, perform probability scale self-organization processing on the feature vectors, and obtain the maximum probability value (PV) generated by the probability scale self-organization max1～n ), the scale of the maximum probability space (PS max1～n ) is the variance of the maximum probability space, and is logged into the database. Then, a face image of a logged-in photo is taken with a camera, and a feature vector (sp_PV) is extracted using the above-mentioned image transformation method or the method of facial features position, or using two or more methods at the same time. max1～n ) to perform face recognition.

[0136] Here, a method for reverse recognition of images is proposed.

[0137] Fig.12 It is a schematic diagram of image recognition of a single photo under multi-scene shooting.

[0138] like Fig.12 As shown in the figure, for each person in the database of more than 10,000 face images, a feature vector (PV) is extracted using the above-mentioned image transformation method or the method of facial features position, or using two methods at the same time, or more methods. Max1～n), and log the feature vector. Then identify a real person's face image from more than 10,000 face images in the database, use a camera to continuously shoot the face of the person to be identified, and obtain several face images of the person to be identified. Then use the above-mentioned image transformation method or the method of facial features position, or use two methods at the same time, or more methods to extract a set of feature vectors, perform probability scale self-organization processing on the feature vectors, and obtain the maximum probability value (sp_PV) generated by probability scale self-organization max1～n ), that is, the expected value of the maximum probability space in statistics, the scale of the maximum probability space (sp_PS max1～n ) is the variance of the maximum probability space, and face recognition is performed. This method can also be called the reverse recognition method.

[0139] Here, a method for performing bidirectional recognition on images is proposed.

[0140] Fig.13 Schematic diagram of single photo image recognition for multi-scene shooting and simulated multi-scene shooting

[0141] like Fig.13 As shown; a camera is used to shoot several faces of the identified persons at the same time or in different scenes to form several face images of the identified persons, or more than 10,000 face images in the database are used, one image for each person, and the above-mentioned image processing method is used to simulate images shot in multiple scenes to obtain several images simulating different scenes, and then the above-mentioned image transformation method or the method of the position of the facial features is used, or two methods or more methods are used at the same time to extract a group of feature vectors, and the probability scale self-organization processing of the feature vectors is performed to obtain the maximum probability value (PV) generated by the probability scale self-organization. max1～n ) That is, the size of the maximum probability space (PS max1～n ) is the maximum probability variance logged into the database.

[0142] Then, a real person is identified from more than 10,000 face images in the database, and the face of the person to be identified is continuously photographed with a camera to obtain a number of face images of the person to be identified. Then, a set of feature vectors is extracted using the above-mentioned image transformation method or the method of facial features position, or two or more methods are used at the same time, and the probability scale self-organization processing of the feature vector is performed to obtain the maximum probability value (sp_PV) generated by the probability scale self-organization max1～n ), the size of the maximum probability space (sp_PS max1～n ) is the maximum probability variance, and face recognition is performed. This method can also be called a two-way recognition method.

[0143] In the following, for quantitative pattern recognition, a method is proposed to segment the probability distribution of the eigenvalues ​​of the logged feature vector.

[0144] Fig.14 It is a schematic diagram of dividing a probability space into multiple probability spaces.

[0145] like Fig.14 As shown; Fig.14 (a) is a complete probability space Φ of eigenvalues, which can be scaled with the maximum probability space according to σ' 3 ~σ' 2 ,σ' 2 ~-σ' 1 ,σ' 1 ~σ 1 , σ 1 ~σ 2 , σ 2 ~σ 3 , divided into Fig.14 (b)

[0146] Φ 1 , Φ 2 , Φ 3 , Φ 4 and Φ 5 Each of the five probability spaces can generate μ through the processing of probability scale self-organizing machine learning. 1 , μ 2 , μ 3 ,

[0147] μ 4 and μ 5 Five maximum probability values, and the scale of the maximum probability space.

[0148] Here, Φ 3 The probability distribution value of the probability space is 68.2%, Φ 2 With Φ 4 The probability distribution value of the probability space is 27.2%, Φ 1 With Φ 5 The probability distribution value of the probability space is 4.3%.

[0149] exist Fig.14 In a complete probability space of (a), if the probability distribution value is only 4.3% of the recognition results appear, it will also be identified as belonging to Fig.14 (a) The probability space corresponding to the recognition result, it is easy to make mistakes in the recognition result. The probability distribution of a feature value is divided into 5 probability distributions. During recognition, if the sample feature value falls in Φ 1 With Φ 5 The probability space can reject the generation of recognition results, and through re-recognition, a recognition result with a higher probability value can be obtained.

[0150] For the probability distribution value of the recognition result of a feature vector, each feature value belonging to the sample feature vector can be placed on the probability distribution value of the evaluation result of the reliability or quality of the feature vector, and then the value can be processed by the machine learning of the probability scale self-organization to obtain the probability value of the recognition result of the feature vector with the maximum probability. According to the probability distribution value of the maximum probability space where the sample data of the recognition result falls, the probability distribution value of the recognition result can be obtained, and it can be judged whether to reject the recognition according to the large or small probability distribution value.

[0151] Next, a method for calculating the accuracy of the recognition result is proposed for the learned and logged feature vector data.

[0152] Assume that there are m eigenvectors in the database, each of which has n eigenvalues. After t training times, the data obtained can be used to obtain the following maximum probability matrix P by using probability scale self-organizing machine learning:

[0153] (Formula 6)

[0154]

[0155] And the scale matrix M of the maximum probability space:

[0156] (Formula 7)

[0157]

[0158] The above two matrices can be analyzed using the (EPD distance) formula that can unify the distance scales of Euclidean space and probability space.

[0159] Fig.15 is a schematic diagram of two probability spaces in Euclidean space.

[0160] like Fig.15 As shown: (∈) is a Euclidean space covering two probability spaces. There are two probability spaces in Euclidean space as well as (ρ 1 ) is the probability space The maximum probability value of 2 ) is the probability space The maximum probability value of . In addition, (σ 1 ) is the probability space The scale of the maximum probability space, (σ 2 ) is the probability space The scale of the maximum probability space, the probability space The probability value of the first maximum probability space Probability Space The probability value of the second maximum probability space Probability Space The probability value of the third maximum probability space Probability Space The probability value of the first maximum probability space Probability Space The probability value of the second maximum probability space Probability Space The probability value of the third maximum probability space

[0161] Furthermore, suppose the probability space as well as The maximum probability value is (ρ 1 ) and (ρ 2 ) are the elements w of the two data sets j ∈W and v j ∈V. Let the maximum probability value be (ρ 1 ) and (ρ 2 ) is connected by a straight line, and there is a sample data point β in the middle of the straight line j ∈B, find the sample data set β j ∈B is the data set to which W or V belongs, which can be obtained by the following EPD distance formula:

[0162] Formula 8

[0163]

[0164]

[0165]

[0166]

[0167] The distance between the unified Euclidean space and the probability space mentioned above satisfies the following conditions:

[0168] (1) Non-negativity:

[0169] (2) Non-degeneracy: d(w, v) = 0, then w = v;

[0170] (3) Symmetry:

[0171] (4) Triangle inequality:

[0172] In order to improve the accuracy of face recognition, a more accurate scale can be introduced that can fully take into account fuzzy information and probability information. A definition method for the probability measure of fuzzy events across different spaces is characterized by: while considering the spatial distance between data, the probability value of the probability distribution of the data in the probability space must also be considered.

[0173] Assume that the set W belongs to the probability space. The formula of the fuzzy event probability measure (FPM measure) of the set V can be obtained by the following formula:

[0174] Formula 9

[0175]

[0176]

[0177]

[0178] Here, pf j (vj) Set element w j ∈W(j=1,2,...,n) is in the set element v j ∈V(j=1, 2, ..., n) probability distribution value of the position in the probability space;

[0179] Similarly, Pf j (wj) is the set element v j ∈V(j=1,2,...,n) is in the set element w j The probability distribution value of the position in the probability space of ∈W(j=1, 2, ..., n).

[0180] d max Yes j to v j The maximum distance.

[0181] The formula of the membership function of the above-mentioned fuzzy event probability measure may include any fuzzy numerical value that can form the objective function according to certain artificially given rules and the result is 0-1. There are various methods for defining similar official script functions, and all methods involving the definition method of fuzzy event probability measure belong to the scope of the present invention.

[0182] The spatial information of the above fuzzy event probability measure is based on the distance between the unified Euclidean space and the probability space, and satisfies the following distance conditions;

[0183] (1) Non-negativity:

[0184] (2) Non-degeneracy: F(w, v) = 0, then w = v;

[0185] (3) Symmetry:

[0186] (4) Triangle inequality:

[0187] Here, in order to improve the accuracy of face recognition, a method is proposed that considers the ratio between the minimum distance between feature vectors and the cumulative value of the scale of the maximum probability space of each eigenvalue of the feature vector.

[0188] Using the above EPD distance formula, we can get the distance matrix between feature vectors:

[0189] (Formula 11)

[0190]

[0191] Here, let d ij =d ji , d ii =d ji =0.

[0192] Using the above FPM measure formula, we can get the measure matrix between eigenvectors:

[0193] (Formula 12)

[0194]

[0195] Here, let d ii =d ji , d ii =d ji = 0, and f ii =f,f ii =f ii =0.

[0196] Here, two eigenvectors corresponding to the minimum EPD distance can be obtained, or two eigenvectors corresponding to the maximum FPM measure can be obtained.

[0197] Then calculate the two closest EPD distance values ​​D min , and the maximum probability space scale of the two eigenvectors corresponding to the two closest EPD distance values and The cumulative value of:

[0198] (Formula 13)

[0199]

[0200] According to formula 10 and formula 12, it can be concluded that the necessary and sufficient conditions for face recognition to achieve the highest recognition accuracy are:

[0201] (Formula 14)

[0202] D min ≥M min or

[0203] Here we can quantitatively calculate the accuracy of face recognition through the data set of the logged feature vector. With such a result, if we need to improve the accuracy of face recognition, we must first solve the quality of feature information extraction of face images. The higher the quality, the better. min With M min The larger the ratio is, the smaller it is or even less than 1, the worse the recognition accuracy is.

[0204] Here, a method is proposed that takes into account the effect of a large-scale registered feature set on the recognition result, and an average distance method that can increase the quality of a feature vector is proposed.

[0205] To increase the quality of the feature vector, that is, to increase the value of (Formula 14), on the one hand, it is necessary to select the image transformation algorithm to be used, and adopt some image transformation algorithms that can achieve higher quality of the feature vector. On the other hand, after determining an image transformation algorithm, the average distance method is adopted to perform a machine learning process of probability scale self-organization on the set of scales of the maximum probability space of each eigenvalue contained in the feature vector. The scale of the maximum probability space of a reliable and high-quality eigenvalue obtained by the machine learning process of probability scale self-organization is used as a benchmark. The eigenvalues ​​are filtered according to the scale and the ones that meet the scale are retained and the ones that do not meet the scale are eliminated. A high-reliability and high-quality eigenvalue set can be obtained. This set is used as the eigenvector for the final comparison of face recognition, and which eigenvector is a high-reliability and high-quality eigenvalue and which is not is represented by a binary code, "1" is a high-reliability and high-quality eigenvalue, and "0" is a discarded eigenvalue, and is stored in the database. In addition, since the number of high-reliability and high-quality eigenvalues ​​obtained by each eigenvector is different, the average distance method is used to unify the scales of each eigenvector.

[0206] In order to improve the accuracy of face recognition, a method is proposed here that considers the common probability distribution information of complex eigenvalues, as well as a method that considers the common probability distribution information of sample eigenvalues ​​at the common position of the common probability distribution of the logged-in complex eigenvalues ​​and the probability distribution of the logged-in eigenvalues.

[0207] Fig.16 It is a schematic diagram of a pattern recognition method that utilizes common probability distribution information.

[0208] like Fig.16 As shown; in pattern recognition, it is often necessary to calculate the two maximum probability eigenvalues ​​ω 4and ω 8 The closest distance to the sample eigenvalue τ can be directly calculated using the above EPD distance formula and FPM measurement formula. 4 When measuring the EPD distance or FPM with τ, ω is not taken into account. 4 and ω 8 When training data here, we can get the matrix of the common maximum probability values ​​between the elements of the eigenvector:

[0209] (Formula 15)

[0210]

[0211] The matrix of the scale of the common maximum probability space between the elements of the eigenvector can also be obtained:

[0212] (Formula 16)

[0213]

[0214] While training the maximum probability value of the feature vector and the scale of the maximum probability space, we also train the common maximum probability value of the two feature vectors and the scale of the common maximum probability space, such as Fig.16 As shown; let φ 1 and φ 3 is the probability distribution of any independent eigenvector among any two different eigenvectors, φ 2 Yes 1 and φ 3 The common probability distribution of 4 is the probability distribution φ 1 The position corresponding to the maximum probability value, ω 8 is the probability distribution φ 3 The position corresponding to the maximum probability value, ω 6 Yes 1 and φ 3 The common probability distribution φ 2 The position corresponding to the maximum probability value of can be obtained by (Formula 8) 1 and φ 3 The corresponding sample eigenvalue τ, the result is from the position of τ to ω 4 The EPD distance of the position is ε 1 , from the position of τ to ω 6 The EPD distance of the position is ε 2 , then the common probability distribution information of the eigenvalues ​​from the position of τ to ω is considered 4 The new distance formula for the position is:

[0215] (Formula 17)

[0216]

[0217] Also considering the common probability distribution information of the eigenvalues ​​from the position of τ to ω 8 The new distance formula for the position is:

[0218] (Formula 18)

[0219]

[0220] Here, we can also assume that the sample image is also a plurality of images read from different environments, and each eigenvalue of the sample feature vector can also generate a probability distribution. In this case, the probability distribution of the eigenvalue of the sample feature vector can be regarded as φ 3 , to solve φ 1 With φ 3 Does it belong to a probability distribution? Using φ 1 With φ 3 The common probability distribution information φ 2 It is easier to get the best recognition result. 1 =ω 6 -ω 4 , ε 2 =ω 8 -ω 6 , and ε 3 =ω 8 -ω 4 ,but

[0221] (Formula 19)

[0222]

[0223] The larger the value of φ 1 With φ 3 The closer the relationship between them.

[0224] Assume that the position of the sample eigenvalue τ is in the probability distribution of the eigenvector φ 1 The probability distribution value of 1 , the position of the sample eigenvalue τ is in the common probability distribution φ of the two eigenvectors 2 The probability distribution value of 2 Similarly, the probability distribution of the position of the sample eigenvalue τ in the eigenvector is φ 3 The probability distribution value of 3 ,but

[0225] After introducing the common probability distribution information of the two eigenvalues, the probability distribution φ of the eigenvector where the sample eigenvalue τ is located 1 The probability distribution value at the position should be:

[0226] (Formula 20)

[0227] P(φ 1 )=P(δ 1 ∩δ 2 )=P(δ 1 )P(δ 2 / δ 1 )

[0228] After introducing the common probability distribution information of the two eigenvalues, the probability distribution of the eigenvector where the sample eigenvalue τ is located is φ 3 The probability distribution value at the position should be:

[0229] (Formula 21)

[0230] P(φ 3 )=P(δ 3 ∩δ 2 )=P(δ 3 )P(δ 2 / δ 3 )

[0231] After introducing the common probability distribution information of the two eigenvalues, τ and ω 4 The new fuzzy event probability measure formula between is:

[0232] (Formula 22)

[0233]

[0234] After introducing the common probability distribution information of the two eigenvalues, τ and ω 8 The new fuzzy event probability measure formula between is:

[0235] (Formula 23)

[0236]

[0237] In order to improve the accuracy of face recognition, the probability information of the fuzzy event probability measure is considered in addition to the sample eigenvalue s j Falling to the eigenvalue f corresponding to a certain login eigenvector j The probability distribution area of ​​the sample eigenvector is the probability distribution value of the area. Another thing is to introduce the common probability distribution information of the two eigenvalues. More importantly, consider the eigenvalues ​​s of the sample eigenvector. j (j=1, 2, ..., n), each eigenvalue f of the corresponding eigenvector falls into j The probability distribution value on the probability distribution area of ​​(j=1, 2, ..., n), or the probability value obtained after introducing the common probability distribution information of each eigenvalue of the two eigenvectors.

[0238] Here, we propose to consider the eigenvalue s of the sample eigenvector S j ∈S(j=1,2,…,n) belongs to a registered feature vector F i (w) The corresponding eigenvalue f ij (w) A method for defining probability information of the maximum probability.

[0239] Fig.17 It is a schematic diagram of a method for defining a new probability value taking into account the overall probability.

[0240] Let the eigenvalue s of the sample eigenvector S be j ∈S(j=1,2,…,n), respectively, in a certain registered feature vector W i The corresponding eigenvalue w ij The probability value of the position on the probability distribution is pf ij (j = 1, 2, ..., n), through the probability scale self-organization pf ij (j=1, 2, ..., n) performs machine learning and obtains a maximum probability value pf i_max (w) , take this probability value as the sample feature value s j With the eigenvalue f ij (w) The probability value of the membership degree considers the relationship between the two feature vectors as a whole, so the recognition accuracy of the image can be improved.

[0241] From this, we can get the sample feature vector S, which belongs to a certain registered feature vector W i The probability measure of fuzzy events is:

[0242] (Formula 24)

[0243]

[0244]

[0245]

[0246] Here, σ 1 Yes ij The maximum probability scale of the probability space, σ 2 Yes ii The scale of the maximum probability of the probability space. From (Formula 23), it can be seen that if the feature vector S→W i FPM i (w) →100, otherwise FPM i (w) →0.

[0247] Similarly, let the eigenvalue s of the sample eigenvector S be j ∈S(j=1,2,…,n), respectively, in a certain registered feature vector V i The corresponding eigenvalues ​​v ij The probability value of the position on the probability distribution is pf ij (v) (j = 1, 2, ..., n), through the probability scale self-organization pf ij (v) (j=1, 2, ..., n) performs machine learning and obtains a maximum probability value pf i_max (v) You can also use pf i_max (v) The value is taken as the sample feature value s j ∈S(j=1,2,…,n) belongs to all eigenvalues ​​v ij The probability value of .

[0248] According to formula 23, the sample feature vector S belongs to a certain registered feature vector V i The probability measure of fuzzy events is:

[0249] (Formula 25)

[0250]

[0251]

[0252]

[0253] From (Formula 24), it can be seen that if S→V i FPM i (v) →100, otherwise FPM i (v) →0.

[0254] By using (Formula 17), (Formula 18) and (Formula 19), s i -v ii ,s i -W ii and w ii -v ij , can be transformed into the distance that takes into account the common probability distribution information of the eigenvalues. As well as the definition method of the probability information of (Formula 20) and (Formula 21), bring it into the above distance formula to obtain the solution of the fuzzy event probability measure. It is also possible to directly use (Formula 22) and (Formula 23) to directly construct the solution of the fuzzy event probability measure.

[0255] The size of the feature vector set should be large. The number of images for face recognition increases. For example, the face is divided into several regions according to the location of the facial features. Each additional region will increase the size of the feature vector exponentially. The increase and decrease of the feature vector data set using the SDL model has a complexity of 0(n 2 ) has a nearly linear characteristic and can increase the size of the feature vector dataset infinitely.

[0256] There are two ways to increase the feature vector data set infinitely; one is Figure 6 As shown in the figure, the image is mapped to a larger feature data set by using the image transformation method. On the other hand, the image is infinitely segmented, and each segment corresponds to an image transformation, thereby obtaining a huge feature data set, and the result can obtain the highest recognition result. This method is not only suitable for face recognition and image recognition, but also for voice recognition and other applications.

[0257] Here we give a machine learning model of probability scale self-organization, and the given probability distribution sequence p 1 , p 2 , ...p m p i ∈G, the element (i = 1, 2, ..., m) of the set, its maximum probability value is A(G), and the scale of the maximum probability space of the maximum probability value A(G) is M[G, A(G)], then, according to the n-1th self-organizing iterative process, the maximum probability value A(G) in the calculated (n-1) )'s maximum probability space scale M[G (n-1 ), A(G (n-1) )] if there are k probability distribution series p (n) 1 , p (n) 2 , ...p (n) k , p (n) i ∈G (n) (i=1, 2, ..., k), then the result of the self-organization of the nth probability scale is:

[0258] (Formula 26)

[0259] A (n) =A(G (n) )

[0260] M (n) =M[G (n) , A(G (n) )]

[0261] G (n) =G{A(G (n-1) ), M[G(n-1) , A(G (n-1) )]}

[0262] The iterative results of the above probability scale self-organization machine learning model can be used for a given sequence p containing random information. 1 , p 2 , ...p m , after iterative processing, we can obtain a maximum probability value that can represent this set of number series, as well as the maximum probability space that exists in this number series, and the scale that can represent the maximum probability of this space.

[0263] When conducting self-organizing machine learning on the probability scale of two-dimensional space or three-dimensional space, it can be achieved by simply replacing the maximum probability value of the two-dimensional space or three-dimensional space with the mechanical center of gravity of the two-dimensional space or three-dimensional space and the scale of the maximum probability space of the two-dimensional space or three-dimensional space with the two-dimensional or three-dimensional variance according to the iterative method mentioned above.

[0264] When performing multi-purpose processing on the objective function, multiple purposes often hope to obtain the best solution in the maximum probability space. In particular, multiple purposes have the multiplication characteristics that can promote the acquisition of the best solution. And often when field feature information is needed, because they are in a common maximum probability space, they can promote each other to automatically generate field features, and can achieve automatic machine learning effects without training. At the same time, in the iterative process of the maximum probability space, it will also promote the convergence of multiple purposes to the best solution. Here, multi-purpose automatic machine learning is proposed.

[0265] Here is an application example of image approximation. The above-mentioned probability scale self-organizing machine learning is introduced. In the image of the maximum probability space for image approximation, the image processing required for image approximation, such as binarization, can be performed. Because the range of the processed image can be narrowed to be closer to the image to be approximated, a more ideal image processing result can be obtained. This is more conducive to the next iteration to obtain an image that is closer to the maximum probability space, and a solution that is closer to the image can be obtained. There is no need to perform prior machine learning on the image to be approximated, and the optimal solution for image approximation can be automatically obtained.

[0266] Fig.18 It is a process flow chart of multi-purpose automatic machine learning.

[0267] like Fig.18 As shown in the flowchart of multi-purpose automatic machine learning, it needs to go through S 1 Initialization step, S 2 Get the maximum probability space step, S 3 Perform the processing steps of purpose 1, S 3+1 Perform the processing steps of purpose 2, ..., S3+n Perform the processing steps for purpose n, S 4 Judgment step, S 5 End step composition.

[0268] Initialization step S 1 : Set the initial conditions for iteration, such as the initial processing method of the image, and the initial maximum probability space scale of the data can be obtained through the probability distribution data of the initial input, or an estimated maximum probability space scale can be directly given. The evaluation function of the iteration effect, in order to prevent the occurrence of infinite iterations, set the maximum number of iterations, etc.

[0269] Get the maximum probability space step S 2 : For a maximum probability space, the scale of the maximum probability space of the space can be obtained. This scale can be used to screen out a space with a larger probability. After obtaining a new maximum probability space, a new scale of the maximum probability space can be generated in the new maximum probability space, which can be used to generate a maximum probability space with a larger probability in the next iteration.

[0270] Perform processing step S of purpose 1 3 : According to the dependency between the processing results of multi-purpose objective functions in the maximum probability space, the objective function of the most prioritized purpose 1 is selected for processing in this step.

[0271] Perform processing step S of purpose 2 3+1 : After performing the processing steps of purpose 1, conditions more suitable for obtaining the best solution are provided for performing the processing steps of purpose 2, so that the objective function of purpose 2 obtains a more prominent best solution. ...

[0273] Perform processing step S for purpose n 3+n : After performing the processing steps of purpose n-1, the most suitable conditions for obtaining the best solution are provided for performing the processing steps of purpose n, so that the objective function of purpose n finally obtains the most outstanding best solution.

[0274] Judgment step 4 :The evaluation function of the iteration effect and the maximum number of iterations are used to judge whether the iteration result is completed. If it is not completed, jump to the maximum probability space step S 2 If completed, go to the next step.

[0275] End step: The program ends, or returns to the main program.

[0276] Radar sensors play an important role in autonomous driving. However, current radar applications simply read radar data to determine obstacles around autonomous vehicles. Since radar information is discrete data, and each data has serious random characteristics, and machines cannot directly understand it, the reliability of obstacle judgment using radar data for autonomous driving is very low.

[0277] Here, a method for processing radar data that can be understood by machines is proposed.

[0278] Fig.19 It is a flow chart of the method for machine understanding radar data processing.

[0279] like Fig.19 As shown; the method for machine understanding radar data processing is divided into: obtaining radar data step S1, performing function interpolation step S2. 2 , perform video transformation step S 3 , and the image recognition step S 4 composed of.

[0280] Step S of obtaining radar data 1 : Get the data of the radar completing a scan of the surrounding obstacles. Generally, 16-line data is composed of 16 lines that can reflect the distance of the surrounding obstacles.

[0281] Perform function interpolation step S 2 :Increasing the number of radar lines can improve the accuracy of observing surrounding obstacles, but the cost increases exponentially with the number of increased lines. How to use lower radar accuracy to obtain higher accuracy application effects? Here, in order to improve the accuracy of radar data and obtain higher application effects with fewer radar lines, radar data can be interpolated by functions. Usually, B-spline function, ordinary spline function, Lagrange interpolation, Hermite interpolation, trigonometric polynomial interpolation, Gaussian trigonometric interpolation, Sink interpolation, multiple regression, least squares method, neighboring point interpolation, bilinear interpolation, bisquare interpolation, bicubic interpolation, cubic interpolation, etc. can be used.

[0282] Perform video transformation step S 3 ,In this step, the discrete radar data is converted into higher ,precision radar data through function interpolation, and then through video transformation, a radar video image with the distance of the ,obstacle as gray value can be obtained.

[0283] Image recognition step S 4: The radar video image obtained in the previous step is the same as the distance video image obtained by the image sensor with the image distance as the grayscale value. Various obstacles are at different distances, so it is very easy to separate various obstacles. Various obstacles are distributed with high density at the corresponding positions in the image space, or can be considered as a high-probability distribution. This distribution can be a three-dimensional cone distribution, or a three-dimensional rectangular distribution, etc. By using this feature and importing the above-mentioned two-dimensional space or three-dimensional space probability scale self-organization, obstacles with three-dimensional cone distribution or three-dimensional rectangular distribution can be directly identified for machine understanding.

[0284] Two images taken at a certain interval by binocular cameras will be misaligned. The size of the misalignment can be used to determine the distance of the object. A monocular camera can also use the constant change of distance when the car is moving. It also uses the information of the misalignment of the images taken before and after to transform the size of the misalignment information into distance information. Therefore, binocular cameras or monocular cameras can also realize grayscale distance images. It is also easy to separate various obstacles, and obstacles are distributed in three-dimensional cones or three-dimensional rectangles in the distance space. Obstacles with three-dimensional cone distribution or three-dimensional rectangular distribution can also be directly identified for machine understanding.

Claims

1. A method for recognizing image depth, characterized in that: In order to achieve image recognition accuracy that exceeds that of traditional machine learning, images can be segmented on a large scale, and then each image segmentation is transformed, thereby obtaining millions or even hundreds of millions of feature sets; Considering the effect of large-scale feature sets on recognition results, an average distance method is proposed to increase the quality of feature vectors. To increase the quality of feature vectors, an image transformation algorithm that can achieve higher quality feature vectors is used. The average distance method is adopted to perform probability scale self-organizing machine learning processing on the set of scales of the maximum probability space of each eigenvalue contained in the feature vector. The scale of the maximum probability space of a reliable and high-quality eigenvalue obtained by the probability scale self-organizing machine learning processing is used as a benchmark. The eigenvalues ​​are screened according to the scale, and those that meet the scale are retained and those that do not meet the scale are eliminated, so that a highly reliable and high-quality set of eigenvalues ​​can be obtained.

Images