Image recognition method

• US 6,236,749 B1
• Filed: 01/29/1999
• Issued: 05/22/2001
• Est. Priority Date: 03/23/1998
• Status: Expired due to Fees

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electronics Corp (Panasonic Holdings Corporation)
• Inventors: Asari, Koji, Satonaka, Takami, Baba, Takaaki
• Primary Examiner(s): Bella, Matthew C.
• Assistant Examiner(s): CHAWAN, SHEELA C
• Application Number:
  US09/147,592
• Time in Patent Office: 844 Days
• Field of Search: 382/156, 382/154, 382/159, 382/160, 382/165, 382/170, 382/168, 382/173, 382/172, 382/181, 382/253, 382/274, 382/276, 382/285, 382/254, 382/250
• US Class Current: 382/154
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 10/751 Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching

Abstract

The method provides object recognition procedure and a neural network by using the discrete-cosine transform (DCT) (4) and histogram adaptive quantization (5). The method employs the DCT transform with the added advantage of having a computationally-efficient and data-independent matrix as an alternative to the Karhunen-Loeve transform or principal component analysis which requires data-independent eigenvectors as a priori information. Since the set of learning samples (1) may be small, we employ a mixture model of prior distributions for accurate estimation of local distribution of feature patterns obtained from several two dimensional images. The model selection method based on the mixture classes is presented to optimize the mixture number and local metric parameters. This method also provides image synthesis to generate a set of image databases to be used for training a neural network.

Claims

1. An image recognition method comprising:

a first step of obtaining quantified image signals by inputting luminance or chromatic signals of all the pixels of a two-dimensional image into a nonlinear table, the two-dimensional image having been obtained by scanning an object in a three-dimensional space; and

a second step of dividing the two-dimensional image, obtained by the first step, into several rectangular elements and obtaining number of pixels belonging to each quantified scale of luminance or chromatic signal in each of the rectangular elements;

wherein the nonlinear table is constructed such that, in histogram distribution of scales which has been obtained by quantizing the input image signals of all the pixels of the two-dimensional image, the histogram of the quantified pixels belonging to each of the scales is equalized;

and wherein a feature pattern of the two-dimensional image obtained from the object in the three-dimensional space is represented by a three-dimensional matrix (x, y, z) by using a pixel number at a horizontal coordinate (x), a pixel number at a vertical coordinate (y) and a coordinate (z) representing an intensity of the quantified image signal within each said rectangular element obtained by performing the first and the second steps.

2. The image recognition method according to claim 1, wherein, instead of obtaining the feature pattern as the three-dimensional matrix (x, y, z), consisting of the pixel number components, from the two-dimensional image by performing the first and the second steps of claim 1 on the image signals of the two-dimensional image which has been obtained from the object in the three-dimensional space,

pixel number components corresponding to a two-dimensional matrix consisting of a horizontal coordinate and a vertical coordinate are taken out from the pixel number components of the three-dimensional matrix with respect to each scale of the intensity of the quantified image signals,

the pixel number components of the two-dimensional matrix are transformed by using a two-dimensional discrete cosine transform or a two-dimensional discrete sine transform, thereby obtaining frequency components in a two-dimensional space, and a feature pattern represented as a three-dimensional matrix (u, v, z) of frequency components, which have been extracted as low-frequency ones from the frequency components in the two-dimensional space such that a number of the components maximizes recognition precision, is used as a feature pattern of the two-dimensional image for the object in the three-dimensional space.

3. The image recognition method according to claim 1, wherein, instead of obtaining the feature pattern as the three-dimensional matrix (x, y, z), consisting of pixel number components, from the two-dimensional image by performing the first and the second steps of claim 1 on the image signals of the two-dimensional image which has been obtained from the object in the three-dimensional space,

the pixel number components of the three-dimensional matrix are transformed by a three-dimensional discrete cosine transform or a three-dimensional discrete sine transform with respect to a horizontal coordinate, a vertical coordinate and a coordinate representing an intensity of the quantified image signal, thereby obtaining frequency components in a three-dimensional space, and

a feature pattern represented as a three-dimensional matrix (u, v, w) of frequency components, which have been extracted as low-frequency ones from the frequency components in the three-dimensional space such that a number of the components maximizes recognition precision, is used as a feature input pattern of the two-dimensional image for the object in the three-dimensional space.

4. The image recognition method according to claim 1,

wherein, in a set of feature patterns, each pattern being represented as the three-dimensional matrix of pixel number components or frequency components of the two-dimensional image, which have been obtained by the method of claim 1 from a set of several two-dimensional image data or from an image database obtained by scanning an object in a three-dimensional space belonging to an arbitrary type of attribute,

a reference vector and a variance vector are obtained from an average and a variance in a distribution of pixel number components or frequency components for each element of each three-dimensional matrix;

the feature pattern of the object in the three-dimensional space belonging to the arbitrary type of attribute is represented by these vectors;

a relative distance between the feature pattern of the three-dimensional matrix, which has been obtained from the two-dimensional image of the object of the arbitrary attribute by the method of claim 1 with reference to the reference vector and the variance vector, and the reference vector is determined; and

the attribute of the object is determined from an arbitrary two-dimensional image other than the image database.

5. The image recognition method according to claim 4,

wherein, in each set of feature patterns having the same attribute of the object in the two-dimensional image database, the database being represented as an image feature pattern of pixel number components or frequency components of the three-dimensional matrix, which has been obtained by the method of claim 1, from a set of several two-dimensional images or the image database obtained from by the object of claim 4 in the three-dimensional space belonging to the arbitrary type of attribute,

a variance vector is obtained for a distribution of pixel number components or frequency components with respect to each element of the three-dimensional matrix by means of a statistical calculation or neural network learning;

and

a reference vector corresponding to an average of the distribution of the pixel number components or the frequency components with respect to each element of the three-dimensional matrix of claim 4 is obtained by means of neural network learning;

the neural network storing the reference vector and the variance vector, which represent the objects having the same attribute of the image database of claim 4, as learning coefficients,

determining a relative distance between the feature pattern of the three-dimensional matrix, which has been obtained from the two-dimensional image other than the image database by the method of claim 1, and the reference vector; and

determining the attribute of the object from an arbitrary two-dimensional image other than the image database.

6. The image recognition method according to claim 4, in each set of feature patterns having the same attribute of the object in the two-dimensional image database, the database being represented as an image feature pattern of pixel number components or frequency components of the three-dimensional matrix, which has been obtained by the method of claim 1 from a set of several two-dimensional images or an image database obtained from by the object of claim 4 in the three-dimensional space belonging to the arbitrary type of attribute,

a variance vector is obtained for a distribution of pixel number components or frequency components with respect to each element of the three-dimensional matrix;

and

a reference value of an amount of mutual information between a component of the variance vector and a component of another variance vector having a different attribute from that of the former variance vector is obtained instead of the variance vector of each said attribute; and

the component of the variance vector of each said attribute is determined by using the components of several variance vectors having respective attributes different from the attribute so as to minimize the reference value of the amount of mutual information.

7. The image recognition method according to claim 5, wherein,

in order to obtain the reference vector and the variance vector, respectively corresponding to the average and the variance of the distribution of pixel number components or frequency components of each said element of the three-dimensional matrix of claim 5, by making a neural network learn,

for the distribution of pixel number components or frequency components of each said element of the three-dimensional matrix which has been obtained by the statistical calculation of claim 5,

not the variance vectors, but a feature pattern, which is separated by a minimum relative distance from a feature pattern belonging to an object having an attribute in the feature pattern space of claim 5 and which has a different attribute from the attribute, is obtained from the feature patterns having the attributes which are represented by the pixel number components or the frequency components of the three-dimensional matrix, obtained by the method of claim 1, for the two-dimensional image in the image database of claim 5,

a feature input pattern, obtained by mixing these feature input patterns with each other, is input to the neural network, and

a mixing ratio thereof is varied in accordance with a number of times of learning.

8. The image recognition method according to claim 5,

wherein in a set of feature patterns, each pattern being represented as the three-dimensional matrix of pixel number components or frequency components of the two-dimensional image, which have been obtained by the method of claim 1 from a set of several two-dimensional image data or from an image database obtained by scanning an object in a three-dimensional space belonging to an arbitrary type of attribute,

a reference pattern vector and a variance pattern vector obtained from an average and a variance in a distribution of feature components for each element of each said three-dimensional matrix are stored by a receiver site of a hardwired network or a wireless network,

the feature input pattern of the two-dimensional image, which is represented by the pixel number components or the frequency components of the three-dimensional matrix obtained by the method of claim 1, is encrypted and encoded by a transmitter site, and

the transmitted feature input pattern of the two-dimensional image is decoded and decrypted by the receiver site, thereby determining the attribute of the object represented by the feature pattern of the two-dimensional image by using the neural network of claim 5.

9. The image recognition method according to one of claims 1, wherein a set of two-dimensional images, which have been obtained by rotating the object in the three-dimensional space several times, each time by a predetermined angle, and by scanning the object every time, is used instead of the set of two-dimensional images or the image database which has been obtained by scanning the object in the three-dimensional space belonging to the attribute of the arbitrary type as described in claim 1, thereby using the feature pattern of pixel number components or frequency components of the three-dimensional matrix of claim 1.