Artificial visual system and method for image recognition

• US 5,675,663 A
• Filed: 03/21/1996
• Issued: 10/07/1997
• Est. Priority Date: 03/22/1995
• Status: Expired due to Fees

• Current Assignee: Honda Giken Kogyo Kabushiki Kaisha (Honda Motor Company)
• Original Assignee: Honda Giken Kogyo Kabushiki Kaisha (Honda Motor Company)
• Inventors: Tsujino, Hiroshi, Koerner, Edgar, Masutani, Tomohiko
• Primary Examiner(s): Cumming, William
• Assistant Examiner(s): Johnson, Brian
• Application Number:
  US08/619,990
• Time in Patent Office: 565 Days
• Field of Search: 382/118, 382/203, 382/216, 382/181, 382/224, 382/226, 382/205, 382/298, 382/299
• US Class Current: 382/181
• CPC Class Codes: G06V 40/164 using holistic features

Abstract

An artificial visual apparatus and method for image recognition having a simple adaptive scaling mechanism enables the definition of scale invariant visual icons in a processing area corresponding to the anterior inferotemporal cortex (AIT) in a one-step, value-based decision making process. Icon related activity states resulting from sensory filtering to a fourth stage KL filter corresponding to the V4 area are recognized independent of the scale and position of the item to be recognized within the maximum visual field. The AIT processing area controls the window of attention in the V4 area and confines further processing onto this selected spotlight. The invention presents a biologically plausible method for scale invariant mapping from the V4 stage filter to the AIT processor. Filtering based on principal component analysis (PCA), or Karhunen-Loeve (KL) filtering, yields image data of the item of interest in the V4 stage filter, such data then being supplied to the AIT processor by a scale-invariant mapping process which controls the number of inputs to the KL filters to achieve constant resolution independent of the scale of the item of interest in the maximum visual field. Thus, the problem of scale-invariant mapping is reduced to a simple adaptive thresholding by feedforward inhibition at the AIT processor.

Claims

1. An artificial visual system comprising:

video camera means for picking up a visual image;

conversion means for converting said visual image into a matrix of pixels, said visual image containing at least one item therein to be recognized by said visual system, said item occupying a portion of said matrix of pixels;

KL (Karhunen-Loeve) filtering means for focusing attention on said item within said matrix of pixels thereby defining said item as a plurality of vectors defining a sub-space of said item within said image, each of said vectors being defined by a predetermined number of KL components;

means for supplying pixel information of said item to a memory means, said means for supplying comprising constant resolution filtering means for limiting the number of KL components per vector in order to maintain a resolution of said item at a predetermined constant resolution;

comparator means for comparing said pixel information supplied from said means for supplying with stored data in said memory means of a similar item in order to recognize said pixel information as corresponding to said item; and

means for triggering formation of an icon representing said item when a result of the comparison indicates an agreement between said pixel information of said item and said stored data.

2. An artificial visual system according to claim 1 wherein said constant resolution filtering means limits the number of KL components according to the following equation (1) such that,

3. An artificial visual system according to claim 1 or claim 2, wherein said constant resolution filtering means limits the number of KL components to a predetermined number M depending on the input, based on the following equation (2), such that, ##EQU5## wherein R is a radius of a window of attention circumscribing said item, F is a maximum frequency of said window of attention, ActPCA(i,F) is an activity (vector value) of an i-th principal component of said frequency F, and C is a constant.

4. An artificial visual system according to claim 1, wherein said KL filtering means comprises a plurality of KL filter systems interconnected together, each KL filter system defining said item as a plurality of vectors defined by a predetermined number of KL components.

5. An artificial visual system according to claim 4, wherein said plurality of KL filter systems comprises first through fourth KL filter systems.

6. An artificial visual system according to claim 5, wherein an output from at least one of said first through fourth KL filter systems is supplied as an input to said memory means.

7. An artificial visual system according to claim 1, further comprising a low pass filter for low pass filtering said image processed by said KL filtering means.

8. An artificial visual system according to claim 1, wherein a pre-attentive processing means defines said visual image containing at least one item therein to be recognized by said visual system.

9. An artificial visual system according to claim 8, wherein said pre-attentive processing means determines a window of attention within said visual image by low pass filtering and/or color filtering.

10. An artificial visual system according to claim 9, further comprising means for supplying a feedback signal from said memory means, following triggering of said icon in said memory means, to said KL filtering means, to enable more refined filtering of said window of attention identified by said pre-attentive processing means.

11. An artificial visual system according to claim 9, further comprising means for supplying a feedback signal from said memory means, following triggering of said icon in said memory means, to a control means for controlling said video camera to thereby cause said video camera to zoom or selectively scan an area within said image identified by the triggered icon so as to optimize a matching of the triggered icon with respective outputs from said KL filtering means.

12. A method for image recognition based on adaptive scaling including the steps of:

picking up a visual image;

converting said visual image into a matrix of pixels, said visual image containing at least one item therein to be recognized, said item occupying a portion of said matrix of pixels;

filtering said image through KL (Karhunen-Loeve) filtering to focus attention on said item within said matrix of pixels, thereby defining said item as a plurality of vectors defining a sub-space of said item within said image, each of said vectors being defined by a predetermined number of KL components;

supplying pixel information of said item to a memory means while limiting the number of KL components per vector in order to maintain a resolution of said item at a predetermined constant resolution;

comparing the supplied pixel information with stored data of a similar item in order to recognize said pixel information as corresponding to said item; and

triggering an icon representing said item when a result of the comparison indicates an agreement between said pixel information of said item and said stored data.

13. A method for image recognition according to claim 12, wherein the constant resolution is maintained according to the following equation (1):

space="preserve" listing-type="equation">RF=C (1)

where R is a radius of a window of attention circumscribing said item, F is a maximum frequency of said window of attention, and C is a constant.

14. A method for image recognition according to claim 12 or 13, wherein the constant resolution is maintained by limiting the number of KL components to a predetermined number M depending on the input according to the following equation (2):

##EQU6## where R is a radius of a window of attention circumscribing said item, F is a maximum frequency of said window of attention, ActPCA(i,F) is an activity (vector value) of an i-th principal component of the frequency F, and C is a constant.

15. A method for image recognition according to claim 12, wherein said step of filtering includes filtering said image through a plurality of interconnected KL filters, each KL filter defining said item as a plurality of vectors each having a predetermined number of KL components.

16. A method for image recognition according to claim 15, wherein said plurality of KL filters comprises first through fourth KL filter systems.

17. A method for image recognition according to claim 12, wherein an output from at least one of said first through fourth KL filter systems is supplied as an input to said memory means.

18. A method for image recognition according to claim 12, further comprising a step of low pass and/or color filtering said image processed in said KL filtering step so as to pre-attentively select a window of attention within said visual image.

19. A method for image recognition according to claim 18, further comprising a step of supplying a feedback signal from said memory means, following triggering of said icon in said memory means, to said KL filtering step, to enable more refined KL filtering of said window of attention.

20. A method for image recognition according to claim 18, further comprising a step of supplying a feedback signal from said memory means, following triggering of said icon in said memory means, to a control means for controlling a video camera that is picking up said visual image to thereby cause said video camera to zoom or selectively scan an area within said image identified by the triggered icon so as to optimize a matching of the triggered icon with respective outputs from said KL filtering step.