Radiant energy – Photocells; circuits and apparatus – Photocell controlled circuit
Reexamination Certificate
1999-03-31
2001-03-06
Allen, Stephone B. (Department: 2878)
Radiant energy
Photocells; circuits and apparatus
Photocell controlled circuit
C250S2140RC
Reexamination Certificate
active
06198089
ABSTRACT:
FIELD OF THE INVENTION
This invention relates to the circuit architecture of two-dimensional steerable and scalable spatial filters for image processing apparatus, and in particular to such filters adapted to spatially filter a set of input currents supplied by an array of photosensitive inputs forming an image sensor.
BACKGROUND OF THE INVENTION
Oriented filters are useful for vision and image processing tasks such as texture analysis, edge detection, and motion analysis. An important sub-class of oriented filters are Gabor filters, which have complex valued convolution kernel
g
(
x
,
y
)
=
1
2
π
σ
ⅇ
-
x
2
+
y
2
2
σ
2
ⅇ
j
(
ω
xo
x
+
ω
yo
y
)
where &sgr;, &ohgr;
xo
, &ohgr;
yo
. This filter is tuned to spatial frequency (&ohgr;
x0
, &ohgr;
y0
) and responds maximally to edges which are oriented at an angle &thgr;=atan(&ohgr;
yo
/&ohgr;
xo
) where &thgr; is defined to be the angle between the horizontal (x) axis and the line perpendicular to the edge. The real part of g(x,y) is often referred to as the even Gabor filter kernel since it is even symmetric. The imaginary part of g(x,y) is often referred to as the odd Gabor filter kernel since it is odd symmetric. Gabor filter kernels have been used to model the receptive fields of neurons in the visual cortex. They have also found applications in computer vision such as stereo vision, binocular vergence control, texture segmentation and face recognition.
One problem with using oriented filters is that it is often necessary to compute the outputs of many filters tuned to different orientations which can be computationally expensive. To overcome this difficulty when using serial digital processing, Freeman and Adelson proposed the use of “steerable filters” (W. T. Freeman & E. H. Adelson, “
The design and use of steerable filters” IEEE Trans. Pattern Analysis and Machine Intelligence,
vol. 13, no. 9, pp 891-906, Sept. 1991) wherein the output of a filter tuned to an arbitrary orientation can be synthesized as a linear combination of the outputs of a small set of “basis filters”. Steerable filters are especially useful for applications where the filter orientation must be adaptively tuned to match the image characteristics. Unfortunately, Gabor filters are not steerable in this sense.
Another important concept in the field of image processing is that of “scale space” (A. P. Witkin “
Scale-space filtering” Proc.
8
th
Int. Joint Conf. Artificial Intelligence,
pp. 1019-1022 IEEE 1983). The concept of scale space is that images consist of details at multiple “scales” or sizes. This may be best understood by reference to an example such as an image of a tree. At the largest scale, the boundary between the tree and the background may be detected, whereas at the smallest scale the boundaries between different leaves may be detected. Scale space filtering refers to the filtering of an image with convolution kernels of different sizes.
PRIOR ART
Several analog circuit architectures and hardware realizations for convolving one dimensional images with Gabor-type filters have been reported. A resistor network with negative resistive connections between second nearest elements was shown to be able to implement an even Gabor-type filter (L. Raffo, S. P. Sabatini, D. D. Caviglia and G. M. Bisio, “
Anisotropic active resistor meshes for implementing image processing operators” Electronics Letters,
vol. 29, no. 22, pp. 960-961 May 27, 1993). Gabor-type filters with arbitrary phase can be obtained by linear combinations of the outputs of three nearest neighbour cells (L. Raffo “
Resistive network implementing maps of Gabor functions of any phase”, Electronics Letters,
vol. 31, no. 22, pp 1913-1914, Oct. 26, 1995), and an implementation of a one dimensional array of nine cells based on transconductance amplifiers is also known (G. M. Bisio, G. M. Bo, M. Confalone, L. Raffo, S. P. Sabatini and M. P. Zizola “
An analog VLSI computational engine for early vision tasks” Proc.
7
th
Intl. Conf. Artificial Neural Networks—ICANN
'97, pp. 1175-80, 1997). Also known are simulations of a double-layer transistor network operating in sub-threshold which overcomes the problem of negative resistance connections between second nearest neighbours (M. Barbaro, A. Nazarro, L. Raffo, “
Synthesis of a recurrent double
-
layer transistor network for early vision tasks” Proc. IEEE International Symposium on Circuits and Systems,
vol. 3, pp. 219-222, May 1998).
The present applicant has previously reported a one dimensional circuit architecture which simultaneously computes both odd and even Gabor-type filter outputs (B. Shi “
Gabor
-
type image filtering with cellular neural networks” Proc. IEEE International Symposium on Circuits and Systems,
vol. 3, pp. 558-561, May 1996). This previous architecture had a number of disadvantages, however. The first disadvantage is that for spatial frequencies less than 2atan0.5, not all filter bandwidths are achievable using only positive resistors. A second disadvantage was that the conductances of the resistors interconnecting the two resistive grids needed to matched with the gains of the transconductance amplifiers. A final disadvantage was that allowing only positive resistance values, the network could only be tuned to positive spatial frequencies. One of the major disadvantages of all of these one dimensional architectures is that it is not clear how they can be extended to two dimensional filtering where the scale and orientation of the Gabor-type filter can be tuned electronically.
SUMMARY OF THE INVENTION
An object of the present invention is to provide an analog circuit architecture which is able to compute the outputs of oriented filters with tunable orientation and scale. The filters are Gabor-type filter, which are similar to Gabor filters except that the functions modulating the complex exponentials are not Gaussian.
According to the present invention therefore there is provided an analog circuit for spatially filtering a plurality of signals simultaneously at inputs to an array comprising:
(a) a first resistive network having a plurality of nodes, each said node being associated with an input to the array,
(b) a second resistive network having a plurality of nodes, each said node being associated with an input to the array,
(c) means for injecting a current into each node of the first resistive network such that the magnitude of said current is dependent upon the voltages and/or currents in the second resistive network associated with nodes corresponding to the nearest neighbours of the corresponding input in the array,
(d) means for injecting a current into each node of the second resistive network such that the magnitude of said current is dependent upon the voltages and/or currents in the first resistive network associated with nodes corresponding to the nearest neighbours of the corresponding input in the array, and
(e) means for providing input to the circuit by injecting a current into each node of the said array such that the magnitude of said current is dependent upon the signal at the input of the array associated with said node.
The nodes are preferably disposed in a two-dimensional array, which is preferably a regular grid such as a rectangular, hexagonal or tridiagonal grid. However, although the invention is particularly useful in the context of a two-dimensional array, it may also be applied to a one-dimensional array which would still have a number of applications such as image tracking and alignment.
In preferred embodiments of the invention the means for injecting current dependent on the signal at each input to the array is a photosensitive element adapted to convert the intensity of light into an electrical current.
Preferably each of the resistive networks comprises transistor circuits such that the resistive connection between any first node and any second node of the circuit comprises a transistor circuit such that the current flowing between the first and second node increases
Allen Stephone B.
Burns Doane Swecker & Mathis L.L.P.
The Hong Kong University of Science and Technology
LandOfFree
Steerable and scalable spatial filters does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Steerable and scalable spatial filters, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Steerable and scalable spatial filters will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2485353