Method for implementing n-dimensional object recognition using dynamic adaptive recognition layers

• US 8,059,890 B2
• Filed: 07/13/2007
• Issued: 11/15/2011
• Est. Priority Date: 03/04/2005
• Status: Expired due to Fees

• Current Assignee: Klaus J. Bach
• Original Assignee: Klaus J. Bach
• Inventors: Bach, Klaus J.
• Primary Examiner(s): DULANEY, KATHLEEN YUAN
• Application Number:
  US11/879,001
• Publication Number:
  US 20070258649A1
• Time in Patent Office: 1,586 Days
• Field of Search: 382/155, 382/190, 382/227
• US Class Current: 382/155
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

In a method and a system for the implementation of multi-layered network object recognition in multi-dimensional space, the structure of a neural recognition network is dynamically generated and adapted to recognize objects. The layers of the network are capable of recognizing key features of the input data by using evaluation rules to establish a hierarchical structure that is independent of data position and orientation, and can adapt varying data densities, geometrical scaling, and faulty or missing data. Adjacent layers of the hierarchy are mutually reinforcing to facilitate the convergence of a solution. Information flow is both bottom-up and top-down during the recognition process providing feedback from higher hierarchical layers to lower layers to cascade the results of higher-level recognition decisions to elements in lower layers.

Claims

1. An object recognition method for implementing object recognition in an N-dimensional space by means of a network having layers of recognition elements (recognition cells contained within the layers, herein called layer cells), the layers being multiple layers, wherein the object recognition method comprises using

a. the network having a multitude of the recognition elements grouped into the layers

i. the layers being organized into a structure which represents a hierarchical order such that layers that are higher in the hierarchical order (higher hierarchical layers) have layer cells connected by links representing ownership to layer cells contained by layers that are lower in the hierarchical order (lower hierarchical layers),

ii. the layers being assigned a multiplicity of certain key features which the layer cells in the assigned layers recognize and represent, with each layer being assigned a different unique group of features to recognize, with the lower hierarchical layers being assigned simple primitive features and the higher hierarchical layers being assigned complex compound features to recognize, with the features assigned to the higher hierarchical layers being more complex assemblages of the simple primitive features assigned to the lower hierarchical layers,

iii. the layers being dynamic in size and in structure in that member layer cells are created or destroyed and links between layer cells of adjacent layers are created or destroyed so as to adapt each higher hierarchical layer to data generated by a next hierarchical layer and recognized by a hierarchical layer,

iv. the layer cells having a polarization vector which indicates a type of feature which each layer cell has recognized if a layer cell represents more than one type of feature and which serves to determine whether said layer cells are compatible with layer cells of neighboring layers, and

v. the layers being further equipped with a rule for determining whether layer cells of subordinate layers should be included in receptive fields of layer cells of higher hierarchical layers,

b. the network being adaptive to the data in that the network represents content data by creating layer cells at locations where input data occur and tolerates an absence of layer cells where there are no data, as a result of a synthesis method which is applied to successive pairs of hierarchically adjacent network layers, where an adjacent layer which is lower in the hierarchical order contains layer cells representing recognition data and an adjacent layer which is higher in the hierarchical order contains no layer cells or insufficient layer cells to account for all of lower layer cells, which are of lower hierarchical layer, the synthesis method comprising the steps of:

i. finding a layer cell on the lower hierarchical layer that is unowned,

ii. generating a new owning layer cell on the higher hierarchical layer and assigning to the new owning layer cell ownership of the lower previously unowned layer cell,

iii. finding all further unowned layer cells on the lower hierarchical layer in a neighborhood of the lower previously unowned layer cell (neighborhood layer cells) and assigning the further unknown layer cells also to the new owning layer cell, such that the new owning layer cell also owns the neighborhood layer cells, thereby forming a new larger group of owned layer cells,

iv. reevaluating or rerecognizing the new owning layer cell based on the new larger group of owned layer cells, and

v. repeating steps i through iv until there are no unowned layer cells on the lower hierarchical layer,

c. the network being further provided with an object recognition method training algorithm in which information flows through the network in both top-down and bottom-up directions to adapt each layer to a presence of layer cells, parameters thereof and a status of polarization vectors (recognition results) of hierarchically adjacent layers and induce convergence to a persistent robust solution, said object recognition method training algorithm comprising adapting, training and converging the network through an iterative process in which cell ownership is modified and layer cells are created or destroyed to converge distribution, parameters and polarization vectors of the layer cells to a final persistent stable state of mutual reinforcement which represents a solution to an object recognition task, said solution being embodied in, and defined by, the layer cells present in the highest layer of the network, said object recognition method being implemented by the steps of

i. Data Input;

copying input data to be recognized into the lowest hierarchical layer of the network,

ii. Initial Construction;

performing the synthesis method to successively construct preliminary elements on each next higher hierarchical layer from a previously-constructed next lower hierarchical layer of the network by means of analytical formulae and/or pattern recognition algorithms to account for all occurrences of patterns on the lower hierarchical layer, until an uppermost level in network hierarchy has been reached and populated with layer cells representing tentative solutions,

iii. Analysis;

for all pairs of cells for which an ownership relationship exists wherein a layer cell in the higher hierarchical layer owns a layer cell in the lower hierarchical layer, successively cascading recognition information from the layer cells of each higher hierarchical layer to owned layer cells in lower hierarchical layers, so as to sharpen and improve a state of accurate lower layer cells and to destroy or weaken poorly defined lower layer cells,

iv. Synthesis;

constructing new and modifying existing elements on each next higher hierarchical layer as was done in the initial construction step (cii) to again create sufficient elements on each layer to own all elements on each next lower hierarchical layer,

v. Iterative Convergence;

repeating the Analysis and Synthesis steps alternatingly until the network has converged and no longer changes, and

vi. Output;

reading the layer cells of the highest layer to an output as recognition results.

2. The object recognition method according to claim 1, wherein the links representing an ownership relationship between layer cells of differing layers are not actually destroyed and created but only deactivated or rendered inactive by setting a weighting factor to zero instead of actually removing the ownership relationship.

3. The object recognition method according to claim 1, wherein the links representing an ownership relationship between layer cells of differing layers are created not just between adjacent layers but also between non-adjacent layers to assist in the object recognition method.

4. The object recognition method according to claim 1, wherein the layers are not arranged linearly, but are arranged in a branched topology.

5. The object recognition method according to claim 1, wherein, for implementing a subset of the object recognition method various steps of the object recognition method are distributed among multiple hard- or software processing systems for improved performance.

6. A system for implementing an object recognition method as claimed in claim 1, comprising a non-transitory computer-readable medium provided with a program for implementing the object recognition method.

7. The system according to claim 6, wherein the network is adaptive to various content data which it represents and which are provided as a result of a creation of layer cells at locations where input data occur, the network however tolerating an absence of layer cells where there are no content data, such adaptivity being effected by a synthesis method which is applied to successive pairs of network layers for any two given adjacent layers where the lower hierarchical layer contains layer cells representing recognition data and the higher hierarchical layer contains no layer cells or insufficient layer cells to account for all lower hierarchical layer cells.

8. The system according to claim 6, wherein the links are provided representing an ownership relationship between layer cells of differing layers, the links being created not just between adjacent layers but also between non-adjacent layers to assist in the object recognition system.

9. The system according to claim 6, wherein the multiple layers are arranged in a branched topology.

10. The system according to claim 6, including an implementation of the object recognition system for distributing the implementation of various tasks among multiple hard- or software processing systems for improved performance.