Method of image analysis

Patent metadata

Field	Value
Publication number	US-9858677-B2
Application number	US-201514844924-A
Country	US
Kind code	B2
Filing date	Sep 3, 2015
Priority date	Sep 5, 2014
Publication date	Jan 2, 2018
Grant date	Jan 2, 2018

Abstract

A method to analyze an image and determine whether to output image data associated with an area of the image is provided. An object detection algorithm using training image data to detect an object based at least in part on a similarity of appearance of image data to data derived from the training image data is provided. Weakly detected objects are classified based on characteristics associated with the weakly detected object and may be added to the training image dataset for use in further training of the object detection algorithm. The object detection algorithm is trained with a revised dataset, the revised dataset being updated with data generated by the object detection algorithm.

First claim

What is claimed is:

1. A method comprising detecting the presence of a first object and a second object in at least one image using an object detection algorithm that uses training image data to train the detection algorithm to detect a given object based at least in part on a similarity of appearance of image data in a corresponding area of a first image of the at least one image to data derived at least in part from the training image data, the method comprising:
providing a first output indicative of a relatively high similarity of appearance of at least part of a first area of the first image to data derived at least in part from the training image data, the first output indicating that a first detection confidence is sufficiently high to indicate that a first object is likely to have been detected in the first area;
determining a characteristic of a second area, of the first image or of a second image of the at least one image, in which an object may be detected, the characteristic being derived from data relating to the first area and being capable of indicating a likelihood of presence of a second object in the second area;
attempting to detect the presence of the second object in the second area;
responsive to a detection of the second object based on a similarity of appearance of at least part of the second area to data derived at least in part from the training image data, determining based on the determined characteristic whether to provide a second output indicating that a second detection confidence is sufficiently high to indicate that the second object is likely to be present in the second area;
outputting image data associated with the second area, the outputted image data being for use in further training of the object detection algorithm;
adding the outputted image data to a dataset that is used to train the object detection algorithm, thereby to generate a revised dataset;
training the object detection algorithm with the revised dataset, thereby to generate a revised trained object detection algorithm; and
testing a performance of the object detection algorithm against that of the revised trained object detection algorithm,
wherein:
the testing is based on determining a ratio of a number of first outputs to a number of second outputs,
the characteristic is proximity to a position,
the number of first outputs is the number of first objects having the first detection confidence and being within a predefined distance of the position, and
the number of second outputs is the number of second objects having the second detection confidence and being within the predefined distance of the position.

2. The method of claim 1 , wherein the characteristic is based at least partly on at least one characteristic selected from the group consisting of:
a proximity of the second area to an estimated position of the second object within an image;
a size of the second object;
a shape of the second object;
a proximity of the first area to the second area;
a relative position of the first area to the second area; and
color information associated with the first object.

3. The method of claim 1 , wherein the testing comprises applying the revised trained object detection algorithm to the dataset and examining separability of positive and negative detections or the number of support vectors for a support vector machine.

4. The method of claim 1 , comprising:
testing a performance of the object detection algorithm against that of the revised trained object detection algorithm; and
if the performance of the revised trained object detection algorithm exceeds that of the object detection algorithm, replacing the object detection algorithm with the revised trained object detection algorithm.

5. The method of claim 4 , further comprising iterating the steps of claim 4 with additional image data until performance of the revised trained object detection algorithm is no longer improved.

6. The method of claim 1 , wherein a strength of similarity is determined by a strength of response of a classifier, the classifier being:
a support vector machine; or
a linear classifier.

7. The method of claim 1 , wherein the indicated likelihood of presence of the second object is weak but not strong, the method further comprising training the object detection algorithm with the revised dataset, thereby to generate, from an object detection database comprising the dataset, a revised trained object detection database comprising the revised dataset.

8. The method of claim 7 , further comprising testing whether performance of the revised trained object detection database is improved relative to that of the object detection database.

9. The method of claim 8 , wherein the testing is based on determining a change in the number of strong detections and the number of weak detections of the object.

10. The method of claim 9 , wherein the image data is added to the dataset based on a determination of a proximity of the weakly but not strongly detected object to a trajectory defined in relation to the image containing the second object.

11. The method of claim 10 , wherein the testing comprises determining a change in density of strongly and weakly detected objects within a predefined distance of the trajectory.

12. A non-transitory computer-readable storage medium comprising computer-executable instructions which, when executed by a processor, causes a computing device to detect the presence of a first object and a second object in at least one image using an object detection algorithm that uses training image data to train the detection algorithm to detect a given object based at least in part on a similarity of appearance of image data in a corresponding area of a first image of the at least one image to data derived at least in part from the training image data, the detecting comprising:
providing a first output indicative of a relatively high similarity of appearance of at least part of a first area of the first image to data derived at least in part from the training image data, the first output indicating that a first detection confidence is sufficiently high to indicate that a first object is likely to have been detected in the first area;
determining a characteristic of a second area, of the first image or of a second image of the at least one image, in which an object may be detected, the characteristic being derived from data relating to the first area and being capable of indicating a likelihood of presence of a second object in the second area;
attempting to detect the presence of the second object in the second area;
responsive to a detection of the second object based on a similarity of appearance of at least part of the second area to data derived at least in part from the training image data, determining based on the determined characteristic whether to provide a second output indicating that a second detection confidence is sufficiently high to indicate that the second object is likely to be present in the second area;
outputting image data associated with the second area, the outputted image data being for use in further training of the object detection algorithm;
adding the outputted image data to a dataset that is used to train the object detection algorithm, thereby to generate a revised dataset;
training the object detection algorithm with the revised dataset, thereby to generate a revised trained object detection algorithm; and
testing a performance of the object detection algorithm against that of the revised trained object detection algorithm,
wherein:
the testing is based on determining a ratio of a number of first outputs to a number of second outputs

Assignees

• Apical Ltd
• Apical Ltd

Inventors

• Romanenko Ilya

Classifications

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06T7/248 Primary

  involving reference images or patches · CPC title

• G06T7/20 Primary

  Analysis of motion (motion estimation for coding, decoding, compressing or decompressing digital video signals H04N19/43, H04N19/51) · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/255

  Detecting or recognising potential candidate objects based on visual cues, e.g. shapes · CPC title