CN101739565A - Large-capacity pattern recognition method - Google Patents

Abstract

The invention provides a large-capacity pattern recognition method, belonging to the technical field of image processing. In the method, weight value matrix elements and the pattern recognition stored on the basis of probability distribution are utilized. The method comprises the steps of: calculating a weight value matrix W according to a provided pattern or an image; designing the matrix elements wi and j of the W to be a random variable or a random numerical value, and forming 2N different Wis according to the number of the elements wi and j to be processed; and finally determining the input pattern to be recognized after training or the image to be the different Wis obtained by measuring and collapsing to achieve the purpose of large-capacity pattern recognition. The method in the invention stores the matrix elements by probability distribution, the recognized pattern or image can reach 2N times of a processing unit, and the storage capacity or the memory capacity is increased by exponential order. The method of the invention is of great significance to the promotion of the researches, such as the pattern recognition, the image processing, human brain consciousness, high-level intelligent robots and the like. The invention is suitable for the recognition of large-capacity pattern or image.

Description

A kind of jumbo mode identification method

Technical field

The present invention relates to high capacity pattern or image efficient identification method, belong to technical field of image processing.

Background technology

The twenties in 20th century is born in pattern-recognition, along with the appearance of the computing machine forties, the fifties artificial intelligence rise, pattern-recognition develops rapidly becomes a subject, theory that it is studied and method have obtained widespread use in a lot of Science and Technology fields, its field relates to artificial intelligence, computer engineering, robot learning, Neurobiology, medical science, the detective learns, aerospace science etc.But in these traditional modes, memory capacity is very limited, and identification all is jumbo image or pattern in actual life, and therefore the realistic problem of many complexity can't solve at all.This just impels the researchist exploring new method always.In recent years, pattern-recognition and other subjects have produced some novel mode identification methods such as combinations such as fuzzy set theory, neural network, quantum calculations, and have become the research focus of pattern-recognition rapidly.

Summary of the invention

The objective of the invention is, design a kind of pattern or image matrix element high capacity highly effective mode identification method, thereby solution traditional mode recognition speed and efficient are not high, be not suitable for discerning the shortcoming of great amount of images or pattern with the probability distribution storage.

Design philosophy of the present invention is:

The present invention uses for reference the core concept of quantum calculation, designs a kind of image matrix element probability storage means, and the memory capacity of this method and recognition efficiency have had exponential raising than classic method.This method has two outstanding characteristics: the matrix of (1) N element can store 2 simultaneously under this method ^NIndividual pattern or image, memory capacity has had exponential raising with respect to traditional storer.This is because an element can be stored two data simultaneously, so a N variable matrix can store 2 ^NIndividual data.(2) owing to the stack of image array, characteristic such as tangle, make calculating can realize concurrent operation, can accelerate computing velocity greatly.This efficient storage transformation has been implemented in storage to image just, has really realized the efficient parallel computing.

Technical scheme of the present invention is a kind of mode identification method that adopts the weight matrix element based on the probability distribution storage.This method is according to the matrix of image state or vector formation, pass through just to develop and determine weights, the design matrix element is with the mode discriminator of probability distribution storage, when external world of input pattern to be identified, through measuring in the memory module or image just can collapse to it with certain probability, so promptly realized the function of pattern-recognition.

Described method is with the matrix element w of W _{I, j}Be designed to a kind of stochastic variable or random number, according to w _{I, j}Numerical values recited on a coordinate axis, be divided into some equal portions, x ₁, x ₂... x _n, the principle of division is to make each matrix element value (having identical element value) will belong in the by stages such as different, subclassification is accurate more more in the interval in theory.If matrix element w _{I, j}Value x _iProbability be ρ _{I, j}, value X so _I+1Probability be 1-ρ _{I, j}ρ wherein _{I, j}=(x _I+1-w _{I, j})/(X _I+1-X _i).

Described method need to determine the w of processing _{I, j}The number of element, be the number N of processing unit and w _{I, j}Value is x _i(probability is ρ _{I, j}) or value be x _I+1(probability is 1-ρ _{I, j}) any arrangement.Because each w _{I, j}Have only two different values, so just can constitute 2 ^NIndividual different W _i, W _iBe the image or the pattern that are stored in the network.

Described method is calculated input pattern and is identified as W _iProbability p _i, according to the character of matrix, can be by probability ρ _{I, j}Constitute probability matrix, its norm of matrix

Sum (N (N-1)) divided by the off-diagonal element of matrix multiply by 2 ^N-1Equal 1.Promptly $\frac{1}{N(N-1)2^{N-1}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|=1$ So $p_i=\underset{i,j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|/N(N-1)2^{N-1}.$

The step of this method of specific implementation is as follows:

(1) with required recognized patterns or image vector orthogonalization: in order to satisfy positivity, the image or the pattern vector that require to provide need be orthogonal vector, because they are not quadratures generally speaking, so must be transformed into orthogonal vector;

(2) the weight matrix W of computation schema or image: calculate weight matrix W according to pattern that provides or image;

(3) the matrix element w that handles according to the feature selecting needs of matrix _{I, j}: we are to the numerical value on the diagonal line among the W and be indifferent to, and W is again symmetric matrix, select matrix element to be processed as required;

(4) with the matrix element w of W _{I, j}Be designed to a kind of stochastic variable or random number: according to w _{I, j}Numerical values recited on a coordinate axis, be divided into some equal portions, with the matrix element w of W _{I, j}Regard a kind of stochastic variable or random number as, i.e. w _{I, j}Both might get its interval left side value, may get interval the right value again with a certain probability;

(5) weight matrix stack expression: weight matrix is expressed as: $W=\underset{i}{\mathrm{\Σ}}{p}_i{W}_i;$

(6) calculating input image is identified as W _iProbability p _i: norm of matrix

Sum (N (N-1)) divided by the off-diagonal element of matrix multiply by 2 ^N-1Equal 1, obtain probability amplitude p respectively _i

(7) pattern-recognition: the image to be identified of training back input is exactly to collapse to different W through measurement _iThereby, reach the purpose of image recognition.

The present invention's beneficial effect compared with the prior art is, the present invention adopts the probability distribution of element among the storage matrix W to construct the linear superposition of memory module, when pattern of input, through relatively will dropping on the linear component of matrix W, thereby reach the purpose of recognition mode with high probability.The pattern or the picture number of conventional store are generally P=0.14N, and N is the number of processing unit, and P is the pattern count of storage.Method among the present invention adopts the storage of matrix element probability distribution, and recognized patterns or image can reach 2 of processing unit ^NDoubly, memory capacity or memory capacity have had exponential raising.Method of the present invention all will have great scientific meaning to the research that promotes pattern-recognition, Flame Image Process, human brain consciousness and high-grade intelligent robot etc.

The present invention is applicable to the identification of high capacity pattern or image.

Description of drawings

Fig. 1 is the process flow diagram of high capacity pattern efficient identification method;

Fig. 2 is according to w _{I, j}Divide the coordinate axis synoptic diagram;

Input pattern or image when Fig. 3 is emulation;

Output mode or image when Fig. 4 is emulation.

Embodiment

The embodiment of the invention is a kind of mode identification method that adopts the weight matrix element based on the probability distribution storage.At first, the element among the weight matrix W that calculates is designed to a kind of stochastic variable or random number,, constructs the linear superposition of memory module on this basis as Fig. 2 according to the pattern or the image calculation weight matrix W that provide.Concrete steps are shown in Fig. 1 process flow diagram.

According to linear superposition, weights can be written as:

$W=\underset{i}{\overset{P_s}{\mathrm{\Σ}}}{p}_i{W}_i---\left(1\right)$

p _iBe that W collapses to W _iProbability (satisfy normalizing condition $\underset{i}{\mathrm{\Σ}}{p}_i=1$ ); P _SBeing the pattern or the total number of images of storage, also is can recognized patterns or total number of images; When external world of input pattern to be identified,, so promptly realized the function of image recognition through measuring in the memory module or image that just can collapse to it with certain probability.

The present invention has designed the performance that an emulation experiment detects the high capacity pattern efficient identification method that proposed.Fig. 3, Fig. 4 are respectively the results of emulation input and output.

This method is on the basis of having analyzed the linear superposition characteristic, the mode discriminator of a kind of storage matrix element based on probability distribution proposed, it has brought up to the processing unit number on memory capacity or memory capacity 2N has doubly had exponential raising than classic method, and its method is:

1.) the weight matrix W of computation schema or image: calculate weight matrix W according to pattern that provides or image.In order to satisfy positivity, the pattern or the image vector that require to provide need be orthogonal vector, because they are not quadratures generally speaking, so must be transformed into orthogonal vector, its conversion can adopt the Gram-Schmidt orthogonalization method to carry out;

2.) with the matrix element w of W _{I, j}Be designed to a kind of stochastic variable or random number: with the matrix element w of W _{I, j}Regard a kind of stochastic variable or random number as, according to w _{I, j}Numerical values recited on a coordinate axis, be divided into some equal portions, x ₁, x ₂... x _nThe principle of dividing is to make each matrix element value (having identical element value) will belong in the by stages such as different, and subclassification is accurate more more in the interval in theory.If matrix element w _{I, j}Value x _iProbability be ρ _{I, j}, value X so _I+1Probability be 1-ρ _{I, j}ρ wherein _{I, j}=(x _I+1-w _{I, j})/(X _I+1-X _i);

3.) handle w as required _{I, j}The number of element, i.e. processing unit number N: the w of Chu Liing as required _{I, j}The number of element, be the number N of processing unit and w _{I, j}Value is x _i(probability is ρ _{I, j}) or value be x _I+1(probability is 1-ρ _{I, j}) any arrangement because each w _{I, j}Have only two different values, so just can constitute 2 ^NIndividual different W _i, W _iBe the pattern or the image that are stored in the network, the image to be identified of input is exactly to collapse to different W through measurement _iThereby, reach the purpose of pattern-recognition.Can discern 2 from the matrix that N element arranged as can be seen here ^NIndividual pattern.The identification capacity, promptly memory capacity has had exponential raising than classic method;

4.) calculating input image is identified as W _iProbability p _i: according to the character of matrix, can be by probability ρ _{I, j}Constitute probability matrix, its norm of matrix

Sum (N (N-1)) divided by the off-diagonal element of matrix multiply by 2 ^N-1Equal 1. promptly

$\frac{1}{N(N-1)2^{N-1}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|=1$

So $p_i=\underset{i,j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|/N(N-1)2^{N-1}.$

Concrete grammar of the present invention is:

The first step: the weight matrix W of computation schema or image

If the vector set of expression pattern has N vectorial V ₁, V ₂... V _nIn order to satisfy positivity, pattern that requirement provides or image vector need be orthogonal vector, because they are not quadratures generally speaking, so must be transformed into orthogonal vector, its conversion can be adopted the Gram-Schmidt orthogonalization method to carry out weight matrix and can be drawn by (1) formula:

$W=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{V}_i{V}_i^T$

Regard the element in the matrix W as stochastic variable, we are to the numerical value on the diagonal line among the W and be indifferent to, and because W is a symmetric matrix, therefore need determine on coordinate axis that the element of its position is less than the weight matrix element far away.

Second step: with the matrix element w of W _{I, j}Be designed to a kind of stochastic variable or random number

Matrix element w with W _{I, j}Regard a kind of stochastic variable or random number as, promptly according to w _{I, j}Numerical values recited on a coordinate axis, be divided into some equal portions x ₁, x ₂... behind the x, w _{I, j}Both might get its interval left side value, may get interval the right value again with a certain probability.The principle that coordinate axis is divided is to make each matrix element value (having identical element value) will belong in the by stages such as different, and subclassification is accurate more more in the interval in theory.If matrix element w _{I, j}Value x _iProbability be ρ _{I, j}, value X so _I+1Probability be 1-ρ _{I, j}, ρ wherein _{I, j}=(x _I+1-w _{I, j})/(X _I+1-X _i).Each w like this _{I, j}Two different values are arranged, determine w to be processed as required _{I, j}After the number, just can constitute 2 ^NIndividual different W _i, W _iBe pattern or image

The 3rd step: calculating input image is identified as W _iProbability p _i

In second step, need to determine processing element number N as required, and constituted 2 ^NIndividual different W _i, according to the character of matrix, can be by probability ρ _{I, j}Constitute probability matrix, its norm of matrix

Sum (N-(N-1)) divided by the off-diagonal element of matrix multiply by 2 ^N-1Equal 1. promptly

$\frac{1}{N(N-1)2^{N-1}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|=1$

So

$p_i=\underset{i,j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|/N(N-1)2^{N-1}$

The i on equal sign both sides is not identical variable

So by (2) Shi Kede:

$W=\underset{i}{\mathrm{\Σ}}{p}_i{W}_i=\underset{i}{\mathrm{\Σ}}\left(\underset{i,j=1}{\overset{N}{\mathrm{\Σ}}}\right|{\mathrm{\ρ}}_{i,j}|/N(N-1)2^{N-1})W_i$

The pattern to be identified of training back input is exactly through measuring with Probability p _iCollapse to different W _iThereby, reach the purpose of pattern-recognition.

Claims (3)

1. the mode identification method of a high capacity highly effective rate is characterized in that, described method adopts the pattern-recognition of weight matrix element based on the probability distribution storage, and described method is with the matrix element w of W _{I, j}Be designed to a kind of stochastic variable or random number, according to w _{I, j}Numerical values recited on a coordinate axis, be divided into some equal portions, X ₁, X ₂... X _n, the principle of division is to make each matrix element value (having identical element value) will belong in the by stages such as different, establishes matrix element w _{I, j}Value X _iProbability be ρ _{I, j}, value X so _J+1Probability be 1-ρ _{I, j}, ρ wherein _{I, j}=(X _I+1-w _{I, j})/(X _I+1-X _i).

2. the mode identification method of a kind of high capacity highly effective rate according to claim 1 is characterized in that, described method need to determine the w of processing _{I, j}The number of element, be the number N of processing unit and w _{I, j}Value is x _i(probability is ρ _{I, j}) or value be x _I+1(probability is 1-ρ _{I, j}) any arrangement because each w _{I, j}Have only two different values, so just can constitute 2 ^NIndividual different W _i, W _iBe the image or the pattern that are stored in the network.

3. according to the mode identification method of the described a kind of high capacity highly effective rate of claim 1, it is characterized in that described method is calculated input pattern and is identified as W _iProbability p _i, according to the character of matrix, can be by probability ρ _{I, j}Constitute probability matrix, its norm of matrix

Sum (N (N-1)) divided by the off-diagonal element of matrix multiply by 2 ^N-1Equal 1, promptly $\frac{1}{N(N-1)2^{N-1}}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|=1$ So $p_i=\underset{i,j=1}{\overset{N}{\mathrm{\Σ}}}\left|{\mathrm{\ρ}}_{i,j}\right|/N(N-1)2^{N-1}.$

Images