COMPRESSIVE SENSING BASED BIO-INSPIRED SHAPE FEATURE DETECTION CMOS IMAGER

• US 20130342681A1
• Filed: 06/24/2013
• Published: 12/26/2013
• Est. Priority Date: 06/22/2012
• Status: Active Grant

• Current Assignee: California Institute of Technology
• Original Assignee: California Institute of Technology
• Inventors: DUONG, Tuan A.
• Granted Patent: US 8,976,269 B2
• Time in Patent Office: Days
• US Class Current: 348/135
• CPC Class Codes: G06V 10/10 Image acquisition document image scanning and transmission H04N1/00; control of digital cameras H04N23/60
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
• CPC Class Codes: G06V 10/449 Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
• CPC Class Codes: G06V 10/955 using specific electronic processors

Abstract

A CMOS imager integrated circuit using compressive sensing and bio-inspired detection is presented which integrates novel functions and algorithms within a novel hardware architecture enabling efficient on-chip implementation.

Claims

1. A compressive sensing-based bio-inspired shape feature detection imager circuit comprising a plurality of circuits operatively coupled to one another, the plurality of circuits comprising:

an active pixel sensor array configured to collect an image in an original space and generate an analog representation of the collected image;

a bio-inspired shape feature compressive sensing projection matrix circuit configured to project the analog representation of the collected image simultaneously onto each target bio-inspired feature of a set of target bio-inspired features and map the projected image from an original space to a compressive sensing space, and generate i) correlation data of the projected image in the compressive sensing space to the set of target bio-inspired features, and ii) reference position data in the original space for the collected image;

a target detection and location circuit configured to process the correlation data and the reference position data to identify a potential target in the collected image from amongst the set of target bio-inspired features;

a compressive sensing sampling data array circuit configured to process the projected image in the compressive sensing space to recover a digital representation of the collected image in the original space, and generate position and identity information of an identified potential target within the recovered collected image in the original space; and

an adaptive target extraction circuit configured to track the identified potential target in a next collected image and extract a corresponding new feature from the next collected image to add to the set of target bio-inspired features.

2. The imager circuit of claim 1, wherein the analog representation of the collected image comprises a plurality of voltages in correspondence of a plurality of pixels of the active pixel sensor array, and wherein the projection matrix circuit is further configured to divide the analog representation of the collected image into a plurality of analog representations in correspondence of a plurality of non-overlapping sub-windows of a same size of the collected image, and project an analog representation of each non-overlapping sub-window sequentially.

3. The imager circuit of claim 2, wherein the projection matrix circuit further comprises a low power hybrid memory-digital-analog-converter (MDAC) circuit and a static random-access memory (SRAM).

4. The imager circuit of claim 3, wherein the set of target bio-inspired features is stored as digital values in the SRAM.

5. The imager circuit of claim 4, wherein for a target bio-inspired feature of the set of target bio-inspired features a plurality of shifted representations across the non-overlapping input sub-windows is stored as digital values in the SRAM.

6. The imager circuit of claim 4, wherein the set of target bio-inspired features is extracted from a statistical data set of a saccadic eye movement samples.

7. The imager circuit of claim 4, wherein the projection matrix circuit is further configured to feed an analog representation of a non-overlapping sub-window to an analog input of the low power MDAC and to initialize a plurality of weights of the low power MDAC with digital values from the SRAM corresponding to the set of target bio-inspired features.

8. The imager circuit of claim 7, wherein the projection matrix circuit performs edge detection on the projected image to map the projected image from the original space to the compressive sensing space.

9. The imager circuit of claim 8, wherein the edge detection uses a principal component analysis technique implemented in the projection matrix circuit.

10. The imager circuit of claim 9, wherein the correlation data are generated using a compressive convolution technique implemented in the projection matrix circuit.

11. The imager circuit of claim 10, wherein the recovery of the collected image is performed by a mutual correlation algorithm implemented in the compressive sensing sampling data array circuit.

12. The imager circuit of claim 11 implemented in an integrated circuit.

13. The imager circuit of claim 12, wherein the integrated circuit is a complementary metal-oxide-semiconductor (CMOS) integrated circuit.

14. A computational camera comprising the imager circuit of claim 11.

15. A method used in an imager integrated circuit (IC) for detecting a potential target feature from a collected image, the method comprising:

storing in the imager IC a plurality of compressive sensing bio-inspired representations of a plurality of target features into a projection matrix;

dividing in the imager IC the collected image into a plurality of non-overlapping sub-windows of a same size;

projecting in the imager IC a sub-window of the plurality of sub-windows onto the projection matrix;

based on the projecting, obtaining in the imager IC a compressive sensing sub-window;

generating in the imager IC correlation data of the compressive sensing sub-window to the plurality of target features;

assigning in the imager IC reference position of the sub-window within the collected image to the correlation data;

repeating the projecting, obtaining, generating and assigning for all sub-windows of the plurality of sub-windows;

evaluating in the imager IC the correlation data for all sub-windows;

based on the evaluating, determining in the imager IC a detected potential target;

recovering in the imager IC the collected image based on the compressive sensing sub-windows; and

based on the assigning and the determining and the recovering, identifying in the imager IC the detected potential target within the recovered collected image.

16. The method of claim 15 further comprising:

collecting in the imager IC a second image;

based on the determining and the assigning, obtaining in the imager IC a new feature of the detected potential target from the second collected image;

storing in the imager IC a compressive sensing bio-inspired representation of the new feature into the projection matrix;

based on the storing, updating in the imager IC the projection matrix; and

using in the imager IC the updated projection matrix for subsequent detecting.

17. The method of claim 16 further comprising removing from the projection matrix in the imager IC one or more representations of one or more target features.

18. The method of claim 17, wherein the projection matrix in the imager IC further comprises a plurality of projection vectors in correspondence of the plurality of compressive sensing bio-inspired representations, and wherein projecting in the imager IC a sub-window of the plurality of sub-windows onto the projection matrix further comprises simultaneously projecting in the imager IC a sub-window of the plurality of sub-windows onto the plurality of projection vectors.

19. The method according to claim 15, wherein the storing further comprises storing in the imager IC for a target feature of the plurality of target features a plurality of compressive sensing bio-inspired representations of a plurality of shifted representations of the target feature across the non-overlapping input sub-windows.