Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
  • G06F17/10—Complex mathematical operations
  • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V10/00—Arrangements for image or video recognition or understanding
  • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
  • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
  • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries

Definitions

• the present invention relates to a method for recognizing objects encoded by vectors. It finds many applications in the field of spect roscopy. More specifically, it can be applied to the Resonance
• Nuclear Magnetic (NMR) mass spectra for the identification of chemical pollutants and gamma ray spectra for the recognition of nuclear waste.
• the method according to the present invention is applicable in neural networks.
• a neural network is a network in which information is processed in parallel. This network seeks to imitate the human brain, in which each neuron functions more or less autonomously.
• a computer neuron processes and stores data independently of other neurons. The organization of such neurons in a network provides a particular global behavior, occasioned by learning, that is to say that each neuron chooses to let, or not to let, pass information according to the information which has already passed through. through the network.
• This learning is characterized, in a computer neural network, by synaptic weights. More mathematically, each neuron sees at its input a sum of information weighted by synaptic weights w.
• a neural network makes it possible to process a lot of information in parallel in order to find a result.
• He is known. in object recognition methods, to use a method called the pseudo-inverse matrix method applied in a neural network.
• This pseudo-inverse matrix method consists notably in determining a GRAM-SCHMIDT [G] matrix from models S, that is to say sets of vectors S ⁇ coding objects to be recognized.
• This GRAM-SCHMIDT matrix is defined by the relation:
• ⁇ and v being numbers of the vectors serving as models for recognition, i indicating the number of the coordinate considered of said vectors, N being the dimension of the representation space considered and P the number of vectors coding the different models of objects to be recognized.
• the synaptic weights w of the neural network can then be determined by the relation:
• the matrix [G] is not invertible, in particular when the coefficients of said matrix are too little different from each other.
• the method described above has a certain number of drawbacks and, in particular, that of not allowing the recognition of objects whose representative vectors taken as references are too similar. There are, for example, many cases of unrecognizable spectra because they are too close.
• the spectra of the gamma rays emitted by this waste are generally very similar to each other. And the method, as explained previously, does not allow the recognition of these gamma ray spectra.
• the object of the present invention is precisely this recognition of very close spectra. It implements a processing of the vectors coding the objects to be recognized before applying the known method of the pseudo-inverse matrix. This processing consists in changing the origin of the representation space in which said vectors are defined.
• the invention relates to a method of recognizing objects coded by vectors in which a neural network is used capable of comparing an unknown vector to reference vectors defined in an original representation space B , each characterized by a label and stored by said neural network, these vectors defining:
• S being a matrix representing said vectors, i indicating the considered coordinate of said vectors, ⁇ (or ⁇ ) indicating the vector considered;
• the pretreatment of the vectors comprises a first step of hierarchical recognition. said first step comprising the following sub-steps:
• One application of the method consists of a signaling device which can be manipulated remotely in order to recognize spectra, in particular of gamma rays emitted by nuclear waste.
• Figures 1a, 1b, 1c, 1d and 1e are the representations of gamma ray spectra from nuclear waste;
• Figure 2 is a schematic representation of a representation space corresponding to Figures 1a to 1e.
• vector will be understood to mean a set of channels describing a spectrum. Mathematically, these vectors are coded by a series of numbers representing the coordinates of said vector in an N-dimensional space and grouped in a rectangular matrix S.
• the values of the coefficients S ⁇ of the GRAM-SCHMIDT matrix [G], of the coefficients d ⁇ of the matrix [d] representing the distances between the vectors S ⁇ and S ⁇ and written: d ⁇ (2 - 2G ⁇ ) 1 ⁇ 2 , and coefficients ⁇ ⁇ of the matrix [ ⁇ ] representing the angular differences between the vectors S ⁇ and S ⁇ and written: ⁇ ⁇ Cos -1 [G ⁇ ], are the following:
• the matrices d and ⁇ are two different ways of expressing the same notion of difference between two vectors S ⁇ and S ⁇ .
• the matrix [G] is a unit matrix, the matrix has all its non-diagonal coefficients d ⁇ equal to and the matrix [ ⁇ ] has, outside its diagonal, only coeffi cient s ⁇ ⁇ of 90 °.
• the matrix [ ⁇ ] confirms what FIGS. 1a and 1e show, namely that the spectra 1 and 5 are very similar to each other: indeed.
• the angle ⁇ 15 between the vectors S 1 and S 2 is only 3 °.
• the angle ⁇ 23 between the vectors S 2 and S 3 is only 7 °, which is very far from the ideal right angle.
• the other angles are of an order of magnitude higher, namely about 25 °.
• This sensitivity index a is defined for an origin B of the representation space.
• it is sought to find a new origin B ′ of the representation space in order to make the vectors S ⁇ and S ⁇ more orthogonal.
• This new origin B ' is found when the sensitivity index ⁇ is minimum.
• the different sensitivity indices ⁇ are chosen using a systematic sampling method from among all the possible origins, any other method of choosing B 'being able to be used.
• the new matrices obtained are the following:
• FIG. 2 is a representation of the representation space of the five vectors of the example considered. Said vectors are represented by points S1, S2, S3, S4 and S5.
• the invention makes it possible to implement a hierarchical preprocessing for vector recognition. It is first of all a question of listing in an increasing order of values all the distances d ⁇ between two vectors S ⁇ and S ⁇ , forming a
• connection matrix [c] the set of determined aggregates: its coefficients C ⁇ where the pair of labels ( ⁇ , ⁇ ) has been taken into account and its coefficients C ⁇ symmetrical of C ⁇ being equal 1; the other coefficients are 0.
• the comparison of an unknown vector is then done in several stages.
• the unknown vector to be determined is first compared to a vector representing the aggregate, said vector being determined as an average of all the vectors contained in an aggregate.
• the unknown vector is then compared to each vector contained in the selected aggregate as being the most similar.
• the list of the distances between the vectors is as follows: d 15 ⁇ d 23 ⁇ d 14 ⁇ d 45 ⁇ d 12 ⁇ d 34 ⁇ d 25 ⁇ d 24 ⁇ d 13 ⁇ d 35
• connection matrix [c] is as follows:
• connection matrix [C '] is then as follows:
• FIG. 2 makes it possible to understand that the new origin B ', chosen without carrying out the hierarchical recognition preprocessing, does not ensure the quasi-orthogonality of the vectors. Indeed, by considering more particularly the aggregate (S 2 , S 3 ), it is seen that the origin B 'chosen, as shown in FIG. 2, does not allow the orthogona li ted between the vectors S 2 and S 3 . On the contrary, if we have chosen the aggregate (S 2 , S 3 ) as the most similar, we can place our in a restricted space with an origin B "allowing a quasi-orthogonality of S 2 and S 3 .
• the invention can indeed be embodied in a manipulative or autonomous signaling device comprising in particular the following organs:
• This mi croca Iculateur has two successive layers of elements:
• the invention also applies to many other fields where the information is codable by a vector, in particular spectroscopy: mass spectra of pollutants, nuclear magnetic resonance spectra, etc.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Theoretical Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Mathematical Physics (AREA)
• Data Mining & Analysis (AREA)
• Computational Mathematics (AREA)
• Mathematical Analysis (AREA)
• Pure & Applied Mathematics (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Computing Systems (AREA)
• Databases & Information Systems (AREA)
• Software Systems (AREA)
• Mathematical Optimization (AREA)
• General Engineering & Computer Science (AREA)
• Algebra (AREA)
• Health & Medical Sciences (AREA)
• Artificial Intelligence (AREA)
• Evolutionary Computation (AREA)
• General Health & Medical Sciences (AREA)
• Medical Informatics (AREA)
• Multimedia (AREA)
• Image Analysis (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=9409064

Family Applications (1)

Country Status (3)

Families Citing this family (6)

• 1991
  • 1991-01-25 FR FR9100867A patent/FR2672141B1/fr not_active Expired - Fee Related
• 1992
  • 1992-01-23 EP EP92904649A patent/EP0568616A1/de not_active Ceased
  • 1992-01-23 WO PCT/FR1992/000060 patent/WO1992013315A1/fr not_active Application Discontinuation

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events