Granted Patent B2

US 12,633,093 · App. 18/179,443 · Granted May 19, 2026

Inventor: Tomoki Taminato (Kanagawa, JP)

Assignee: CANON KABUSHIKI KAISHA

G06V10/776G06V10/7788

Abstract

Claims (74)

1 . An information processing apparatus comprising:

a processor; and

a memory including instructions stored thereon, which when executed by the processor cause the information processing apparatus to:

update a weight using corresponding learning data, with respect to each of a plurality of recognition models;

evaluate recognition accuracy with respect to each of the recognition models;

set parameters for integrating weights corresponding to a common layer of the recognition models based on results of evaluation of the recognition accuracy;

integrate the weights corresponding to the common layer based on the parameters; and

correct the weight corresponding to the common layer based on the recognition accuracy, with respect to each of the recognition models,

wherein the plurality of recognition models are integrated in such a manner that the recognition models share the common layer, by integrating the weights corrected by the first correcting unit,

wherein the weight is corrected corresponding to the common layer based on the highest recognition accuracy in a predetermined period, with respect to each of the recognition models, and

wherein correcting the weight corresponding to the common layer comprises applying a correction defined by W′=(1−β)×W+β×W best , where W is the weight corresponding to the common layer before correction, W best is a weight corresponding to the common layer at a point in time when the recognition accuracy was the highest in a predetermined period, and β is a coefficient predetermined for the correction.

2 . The information processing apparatus according to claim 1 ,

wherein each of the plurality of recognition models performs the same recognition task.

3 . The information processing apparatus according to claim 1 ,

wherein one of the plurality of recognition models performs a different recognition task from another recognition model of the plurality of recognition models.

4 . The information processing apparatus according to claim 1 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

to set a target for recognition accuracy of each of the recognition models,

wherein the recognition accuracy is evaluated based on the target.

5 . The information processing apparatus according to claim 4 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

set the target as a predetermined value or sets the target based on recognition accuracy of the recognition model that has been evaluated before the weight is updated.

6 . The information processing apparatus according to claim 1 ,

wherein the parameters include an integration rate to be used to integrate the weights through weighted averaging.

7 . The information processing apparatus according to claim 1 ,

wherein the parameters include the number of repetitions of updating of the weight to be performed using learning data until integration is performed once.

8 . The information processing apparatus according to claim 6 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

increase the integration rate if the recognition accuracy is degraded compared with that before integration is performed, and maintains or reduces the integration rate if the recognition accuracy is improved compared with that before integration is performed.

9 . The information processing apparatus according to claim 7 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

increase the number of repetitions if the recognition accuracy is degraded compared with that before integration is performed, and maintains or reduces the number of repetitions if the recognition accuracy is improved compared with that before integration is performed.

10 . The information processing apparatus according to claim 1 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

present the recognition accuracy to a user; and

set the parameters by obtaining the parameters that are input by the user.

11 . The information processing apparatus according to claim 1 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

generate a learned model that carries out reinforcement learning to estimate the parameters from the recognition accuracy; and

set the parameters based on the learned model.

12 . The information processing apparatus according to claim 1 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

update the weight corresponding to the common layer of each of the plurality of recognition models by using the weight integrated.

13 . The information processing apparatus according to claim 12 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

determine whether or not the recognition accuracy of the recognition model is degraded as a result of updating of the weight,

wherein the updating is not executed if it is determined that the recognition accuracy is degraded.

14 . The information processing apparatus according to claim 12 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

correct, with respect to each of the recognition models, the weight corresponding to the common layer and updated, based on the recognition accuracy of the recognition model.

15 . The information processing apparatus according to claim 1 , wherein the instructions, when executed by the processor further cause the information processing apparatus to:

set a set to which two or more of the recognition models belong; and

integrate the recognition models belonging to the set.

16 . The information processing apparatus according to claim 1 ,

wherein the weight is updated with respect to each of the plurality of recognition models that respectively run on separate terminals.

17 . The information processing apparatus according to claim 1 ,

wherein the information processing apparatus performs a recognition task with respect to an image that is input, by using a recognition model into which the plurality of recognition models are integrated.

18 . The information processing apparatus according to claim 1 ,

wherein the recognition models perform one or more tasks out of a classification task for identifying an object in an image, an object detection task for detecting the location of an object, a region dividing task for extracting a specific region, a tracking task for tracking a subject, a motion recognition task for identifying motions of a subject, an organ detection task for detecting an organ, including: the face, of a subject, a posture estimation task for estimating a posture of a subject, and a depth estimation task for estimating the depth of an object.

19 . An information processing method comprising:

updating a weight using corresponding learning data, with respect to each of a plurality of recognition models;

evaluating recognition accuracy with respect to each of the recognition models;

setting parameters for integrating weights corresponding to a common layer of the recognition models based on results of evaluation of the recognition accuracy;

integrating the weights corresponding to the common layer based on the parameters; and

correcting the weight corresponding to the common layer based on the recognition accuracy, with respect to each of the recognition models,

wherein the plurality of recognition models is integrated in such a manner that the recognition models share the common layer, by integrating the corrected weights, and

wherein the weight corresponding to the common layer is corrected based on the highest recognition accuracy in a predetermined period, with respect to each of the recognition models, and

wherein correcting the weight corresponding to the common layer comprises applying a correction defined by W′=(1−β)×W+β×W best , where W is the weight corresponding to the common layer before correction, W best is a weight corresponding to the common layer at a point in time when the recognition accuracy was the highest in a predetermined period, and β is a coefficient predetermined for the correction.

20 . A non-transitory computer readable storage medium storing a program that, when executed by a computer, causes the computer to perform an information processing method comprising:

updating a weight using corresponding learning data, with respect to each of a plurality of recognition models;

evaluating recognition accuracy with respect to each of the recognition models;

setting parameters for integrating weights corresponding to a common layer of the recognition models based on results of evaluation of the recognition accuracy;

integrating the weights corresponding to the common layer based on the parameters; and

correcting the weight corresponding to the common layer based on the recognition accuracy, with respect to each of the recognition models,

wherein the plurality of recognition models is integrated in such a manner that the recognition models share the common layer, by integrating the corrected weights,

wherein the weight corresponding to the common layer is corrected based on the highest recognition accuracy in a predetermined period, with respect to each of the recognition models, and

wherein correcting the weight corresponding to the common layer comprises applying a correction defined by W′=(1−β)×W+β×W best , where W is the weight corresponding to the common layer before correction, W best is a weight corresponding to the common layer at a point in time when the recognition accuracy was the highest in a predetermined period, and β is a coefficient predetermined for the correction.

21 . An integrated model comprising:

a recognition model comprising: a common layer; and one or more model-specific layers configured to perform recognition in cooperation with the common layer; and

at least one other recognition model,

wherein the recognition model and the at least one other recognition model are integrated such that the common layer is shared,

wherein a weight corresponding to the common layer is an integrated weight obtained by integrating the weights corresponding to the common layer of each of the recognition models based on parameters set according to the recognition accuracy of each of the recognition models, and

wherein the weight corresponding to the common layer of each of the recognition models is a corrected weight obtained by applying a correction defined by W′=(1−β)×W+β×W best , where W is the weight corresponding to the common layer before correction, W best is a weight corresponding to the common layer at a point in time when the recognition accuracy was highest in a predetermined period, and β is a coefficient predetermined for the correction.

Assignments (1)

Priority Claims (2)

Continuity (1)

References Cited (13)