Filtering methods for visual object detection

Patent metadata

Field	Value
Publication number	US-10169661-B2
Application number	US-201514665687-A
Country	US
Kind code	B2
Filing date	Mar 23, 2015
Priority date	Mar 28, 2014
Publication date	Jan 1, 2019
Grant date	Jan 1, 2019

Abstract

Machine logic that pre-processes and post-processes images for visual object detection by performing the following steps: receiving a set of image(s); filtering the set of image(s) using a set of multimodal integral filter(s), thereby removing at least a portion of the set of image(s) and resulting in a filtered set of image(s); performing object detection on the filtered set of image(s) to generate a set of object-detected image(s); assembling a first plurality of object-detected image(s) from the set of object-detected image(s); and upon assembling the first plurality of object-detected image(s), performing non-maximum suppression on the assembled first plurality of object-detected image(s).

First claim

What is claimed is:

1. A method comprising:
receiving a set of images;
filtering the set of images using a set of integral filters, resulting in a filtered set of images; and
performing machine-learning based object detection on the filtered set of images to generate a set of object-detected images,
wherein filtering the set of images using the set of integral filters includes:
determining a candidate image region of an image of the set of images, wherein the candidate image region corresponds to a particular modality, wherein the particular modality includes one of: (i) color channels, (ii) contextual background selection, and (iii) pixel intensities,
for the image of the set of images, computing an integral image based on the particular modality,
computing integral image output, including at least an average of pixels of the particular modality, utilizing the computed integral image, and
based, at least in part, on the computed integral image output, removing the candidate image region from the image, wherein the image no longer includes the candidate image region.

2. The method of claim 1 , wherein the set of images includes surveillance output collected over a period of time from one or more surveillance devices.

3. The method of claim 1 , wherein the machine-learning based object detection is performed by a face detector, and wherein the method further comprises:
performing face-like background filtering on the filtered set of images, wherein performing the face-like background filtering on the filtered set of images includes:
running the face detector at a sensitive threshold on background regions of filtered images of the filtered set of images, the background regions detected by background modeling,
recording a number of face detections for candidate regions within the background regions over a period of time, and
in response to determining that the number of face detections for a candidate region exceeds a threshold, discarding the candidate region when using the face detector to perform machine-learning based object detection.

4. The method of claim 1 , further comprising:
assembling a first plurality of object-detected images from the set of object-detected images; and
performing non-maximum suppression on the assembled first plurality of object-detected images, wherein the non-maximum suppression is performed on the images of the assembled first plurality of object-detected images simultaneously as a group.

5. A computer program product comprising a computer readable storage medium having stored thereon:
program instructions programmed to receive a set of images;
program instructions programmed to filter the set of images using a set of integral filters, resulting in a filtered set of images; and
program instructions programmed to perform machine-learning based object detection on the filtered set of images to generate a set of object-detected images,
wherein filtering the set of images using the set of integral filters includes:
determining a candidate image region of an image of the set of images, wherein the candidate image region corresponds to a particular modality, wherein the particular modality includes one of: (i) color channels, (ii) contextual background selection, and (iii) pixel intensities, for the image of the set of images, computing an integral image based on the particular modality,
computing integral image output, including at least an average of pixels of the particular modality, utilizing the computed integral image, and
based, at least in part, on the computed integral image output, removing the candidate image region from the image, wherein the image no longer includes the candidate image region.

6. The computer program product of claim 5 , wherein the set of images includes surveillance output collected over a period of time from one or more surveillance devices.

7. The computer program product of claim 5 , wherein the machine-learning based object detection is performed by a face detector, and wherein the computer program product further comprises:
program instructions programmed to perform face-like background filtering on the filtered set of images, wherein performing the face-like background filtering on the filtered set of images includes:
running the face detector at a sensitive threshold on background regions of filtered images of the filtered set of images, the background regions detected by background modeling,
recording a number of face detections for candidate regions within the background regions over a period of time, and
in response to determining that the number of face detections for a candidate region exceeds a threshold, discarding the candidate region when using the face detector to perform machine-learning based object detection.

8. The computer program product of claim 5 , further comprising:
program instructions programmed to assemble a first plurality of object-detected images from the set of object-detected images; and
program instructions programmed to perform non-maximum suppression on the assembled first plurality of object-detected images, wherein the non-maximum suppression is performed on the images of the assembled first plurality of object-detected images simultaneously as a group.

9. The computer program product of claim 8 , further comprising:
program instructions programmed to assemble a second plurality of object-detected images from the set of object-detected images, where the second plurality of object-detected images includes at least one object-detected image from the first plurality of object-detected images and at least one object-detected image that is not included in the first plurality of object-detected images; and
program instructions programmed to perform non-maximum suppression on the assembled second plurality of object-detected images, wherein the non-maximum suppression is performed on the images of the assembled second plurality of object-detected images simultaneously as a group.

10. A computer system comprising:
one or more processors; and
a computer readable storage medium;
wherein:
the one or more processors are programmed to run program instructions stored on the computer readable storage medium; and
the program instructions include:
program instructions programmed to receive a set of images;
program instructions programmed to filter the set of images using a set of integral filters, resulting in a filtered set of images; and
program instructions programmed to perform machine-learning based object detection on the filtered set of images to generate a set of object-detected images,
wherein filtering the set of images using the set of integral filters includes:
determining a candidate image region of an image of the set of images, wherein the candidate image region corresponds to a particular modality, wherein the particular modality includes one of: (i) color channels, (ii) contextual background selection, and (iii) pixel intensities,
for the image of the set of images, computing an integral image based on the particular modality,
computing integral image output, including at least an average of pixels of the particular modality, utilizing the computed integral image, and
based, at least in part, on the computed integral image output, removing the candidate image region from the image, wherein the image no longer includes the candidate image region.

11. The computer system of claim 10 , wherein the set of images includes surveillance output collected over a period of time from one or more surveillance devices.

12. The computer system of claim 10 , wherein the machine-learning based object detection is performed by a face dete

Assignees

• IBM

Inventors

• Bobbitt Russell P
• Brown Lisa M
• Feris Rogerio S
• Pankanti Sharathchandra U
• Zhai Yun

Classifications

• G06V10/56

  relating to colour · CPC title

• G06V10/446

  using Haar-like filters, e.g. using integral image techniques · CPC title

• G06K2209/21

  Physics · mapped topic

• G06K9/00771 Primary

  Physics · mapped topic

• G06K9/4652

  Physics · mapped topic