Neural networks for machine vision

• US 4,803,736 A
• Filed: 07/23/1987
• Issued: 02/07/1989
• Est. Priority Date: 11/27/1985
• Status: Expired due to Term

• Current Assignee: Trustees of Boston University
• Original Assignee: Trustees of Boston University
• Inventors: Mingolla, Ennio, Grossberg, Stephen
• Primary Examiner(s): Blum, Theodore M.
• Application Number:
  US07/102,018
• Time in Patent Office: 565 Days
• Field of Search: 382/10, 382/15, 382/36, 382/22, 382/54
• US Class Current: 382/173
• CPC Class Codes: G06T 7/12 Edge-based segmentation
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]

Abstract

Network interactions within a Boundary Contour (BC) System, a Feature Contour (FC) System, and an Object Recognition (OR) System are employed to provide a computer vision system capable of recognizing emerging segmentations. The BC System is defined by a hierarchy of orientationally tuned interactions, which can be divided into two successive subsystems called the OC filter and the CC loop. The OC filter contains oriented receptive fields or masks, which are sensitive to different properties of image contrasts. The OC filter generates inputs to the CC loop, which contains successive stages of spatially shore-range competitive interactions and spatially long-range cooperative interactions. Feedback between the competitive and cooperative stages synthesizes a global context-sensitive segmentation from among the many possible groupings of local featural elements.

Claims

1. A Boundary Contour System for detecting the boundaries of images having contrast distribution comprising:

(a) mask means responsive to signals corresponding to said contrast distribution for weighting such signals with an orientational weighting function to produce mask output signals proportional to the degree of orientational congruence between the weighting function and the contrast distribution;

(b) node means responsive to said mask output signals for enhancing said mask signals by suppressing the strength of mask output signals emanating from regions of the image of homogeneous contrast distribution to produce enhanced orientational mask signals;

(c) feedback means responsive simultaneously to the orientational signals emanating from either side of spatial domains about a region of said image for generating a feedback signal to said node means when the orientated signals from both sides of said spatial domains are of sufficient value and orientation to indicate that a boundary contour is present in the region of the image between said domains.

2. The system of claim 1 wherein the image is formed of pixels representing light intensities of a visual image.

3. A system for detecting and defining the boundaries of perceptually significant domains in a spatial image having distinct contrast distribution comprising:

(a) an array of families of oriented masks for sampling said image, with each family of masks corresponding to a respective position relative to said image, and with each such mask in a family corresponding to a different orientation relative to said image and each such mask providing a mask output signal proportional to the degree of congruence between the weighting function of the mask function and the distribution of contrasts in the image region sampled by said mask, and wherein said output signal is sensitive to the position, orientation of the contrast and amount of contrast in the regions sampled by the image and insensitive to the direction of contrast in the region sampled by the image;

(b) a first competitive stage comprising a first array of network nodes, one for each mask position and orientation, to which said mask output signals are coupled and wherein mask signals of a given orientation exert an excitatory effect upon nodal signals of similar orientation at a corresponding nodal position of the first array of network nodes and exert an inhibatory effect upon nodal signals of similar orientation at proximate neighboring nodes of the first array;

(c) a second competitive stage comprising a second array of network nodes, one for each node in said first array, to which the respective nodal signals from the first competitive stage are respectively coupled and wherein nodal signals from the first competitive stage exert an inhibitory effect on nodal signals of other dissimilar orientations and exert an excitatory effect on nodal signals of similar orientation at corresponding nodal positions of the second competitive stage;

(d) a cooperative stage comprising a third array of network nodes, one for each node in the second array, each such node of the third array being coupled to a plurality of nodes of the second array within two domains, one on either side of the corresponding node in the second array, and wherein to the extent there is sufficient alignment, as sensed from said signals, of the orientation of the plurality of nodes in both domains, with the orientation of such node of the third array an alignment signal is generated at such node of the third array;

(e) a feedback stage comprising;

a fourth array of network nodes, one for each node in the third array, to which respective alignment signals are coupled and wherein alignment completion signals of a given orientation exert an excitatory effect on nodal signals of similar orientation at a corresponding nodal position of the fourth array of network nodes and exert an inhibitory effect on signals of similar orientation at neighboring nodes of the fourth array; and

(f) a means for coupling the signals from respective nodes of said fourth array to the respective nodes of the first array.

4. A method for detecting and defining the boundaries of perceptually significant domains in a spatial image having distinct contrast distribution comprising:

(a) sampling said image with an array of families of oriented masks, with each mask in a family providing a mask output signal proportional to the degree of congruence between the weighting function of the mask and the distribution of contrasts in the image region sampled by said mask;

(b) coupling said mask signals to respective nodes of a first array of network nodes, one for each mask position and orientation, such that mask signals of a given orientation exert an excitatory effect upon nodal signals of similar orientation at a corresponding nodal position of the first array of network nodes and exert an inhibitory effect upon nodal signals of similar orientation at proximate neighboring nodes of the first array;

(c) coupling the nodal signals of the first array to respective nodes of a second array of network nodes, one for each node in said first array, such that nodal signals from the first array exert an inhibitory effect on nodal signals of dissimilar orientations and exert an excitatory effect on nodal signals of similar orientations at corresponding nodal positions of the second array;

(d) coupling the nodal signals from a plurality of nodes in the second array within two domains on either side of a given node in the second array to a respective node in the third array corresponding to said given node in the second array, and wherein, to the extent there is sufficient alignment of the orientation of the plurality of nodes in both domains, as sensed from said nodal signals, with the orientation of such node of the third array, generating alignment signals from each such node of the third array;

(e) coupling the alignment signals to a fourth array of network nodes, one for each node in the third array, such that signals of a given orientation exert an excitatory effect on nodal signals of similar orientation at a corresponding nodal position of the fourth array of network nodes and exert an inhibitory effect on signals of similar orientation at neighboring nodes of the fourth array; and

(f) coupling the signals from respective nodes of said fourth array to the respective nodes of the first array to provide a feedback signal at each such node of the first array.

5. A Boundary Contour System for detecting the boundaries of images having contrast distribution comprising:

(a) mask means with orientation k and centered at position i, j, responsive to preprocessed signals s_pq corresponding to said contrast distribution for producing an output signal J_ijk equal to ##EQU20## (b) first node means responsive to said mask output signals J_ijk for actuating a first potential W_ijk and inhibiting second potential W_pqk if |p-1|² +|q-j|² is sufficiently small;

(c) second node means responsive to potential W_ijk and a potential perpendicular thereto W_ijK for producing third and fourth potentials X_ijk and X_ijK representing competition between W_ijk and W_ijK at their target potentials;

(d) a shunting network for interacting the first, second, third and fourth potentials at each position i,j, to produce fifth and sixth potentials Y_ijk and Y_ijK which satisfy the equation ##EQU21## wherein D expresses the rate of passive decay of activity; and

E expresses the saturation level of Y_ijk ;

(e) feedback means responsive simultaneously to the said potentials emanating from either side of spatial domains about a region of said image for generating a feedback signal to said first node means when the potentials from both domains are of sufficient value and orientation to indicate that a boundary contour is present in the region of the image between domains.

6. The system of claim 5 wherein the output of the first and second node means is normalized to insure that the potentials resulting from all orientations for a given position is bounded.

7. The system of claim 6 wherein the value of the third and fourth potentials X_ijk and X_ijK is proportional to the ratio of the potential for a given activated orientation versus the potential of the sum of all other actuated orientations at said given orientation.