Domain adaptation for image classification with class priors

Patent metadata

Field	Value
Publication number	US-9710729-B2
Application number	US-201414477215-A
Country	US
Kind code	B2
Filing date	Sep 4, 2014
Priority date	Sep 4, 2014
Publication date	Jul 18, 2017
Grant date	Jul 18, 2017

Abstract

In camera-based object labeling, boost classifier ƒ T (x)=Σ r=1 M β r h r (x) is trained to classify an image represented by feature vector x using a target domain training set D T of labeled feature vectors representing images acquired by the same camera and a plurality of source domain training sets D S 1 , . . . , D S N acquired by other cameras. The training applies an adaptive boosting (AdaBoost) algorithm to generate base classifiers h r (x) and weights β r . The r th iteration of the AdaBoost algorithm trains candidate base classifiers h r k (x) each trained on a training set D T ∪D S k , and selects h r (x) from previously trained candidate base classifiers. The target domain training set D T may be expanded based on a prior estimate of the labels distribution for the target domain. The object labeling system may be a vehicle identification system, a machine vision article inspection system, or so forth.

First claim

The invention claimed is:

1. A labeling system comprising:
an electronic data processing device configured to label an image to be labeled belonging to a target domain by operations including:
training a boost classifier f T (x)=Σ r=1 M β r h r (x) to classify an image belonging to the target domain and represented by a feature vector x in a feature space X, the training using a target domain training set D T comprising labeled feature vectors in the feature space X representing images belonging to the target domain and a plurality of source domain training sets D S 1 , . . . , D S N where N≧2 comprising labeled feature vectors in the feature space X representing images belonging to source domains S 1 , . . . , S N respectively, the training comprising expanding the target domain training set D T based on a prior estimate of the labels distribution for the target domain and, after expanding the target domain training set D T , applying an adaptive boosting (AdaBoost) algorithm to generate the base classifiers h r (x) and the base classifier weights β r of the boost classifier f T (x), wherein the r th iteration of the AdaBoost algorithm includes (i) performing N sub-iterations in which the k th sub-iteration trains a candidate base classifier h r k (x) on a training set combining the target domain training set D T and the source domain training set D S k and (ii) selecting h r (x) as the candidate base classifier with lowest error for the target domain training set D T ;
computing a feature vector x in in the feature space X representing the image to be labeled; and
generating a label for the image to be labeled by operations including evaluating f T (x in )=Σ r=1 M β r h r (x in ).

2. The labeling system of claim 1 wherein the labeling system is a camera-based object labeling system further comprising:
a system camera arranged to acquire images of objects;
wherein the target domain is defined as the domain of images of objects acquired by the system camera and the image to be labeled is an image of an object to be labeled acquired by the system camera;
wherein each source domain S 1 , . . . , S N is defined as the domain of images of objects acquired by a camera other than the system camera; and
wherein the electronic data processing device is further configured to generate a label for the object to be labeled based on the label generated for the image to be labeled.

3. The camera-based object labeling system of claim 2 further comprising:
a display device operatively connected with the electronic data processing device to display the image of the object to be labeled together with the label generated for the object.

4. The labeling system of claim 1 wherein the k th sub-iteration trains the candidate base classifier h r k (x) on a union D T ∪D S k of the target domain training set D T and the source training set D S k .

5. The labeling system of claim 1 wherein the training of the boost classifier f T (x)=Σ r=1 M β r h r (x) further comprises:
before applying the Adaboost algorithm, performing unsupervised source-target domain alignment to align the target domain training set D T and the source training sets D S k , k=1, . . . , N.

6. The labeling system of claim 1 wherein the r th iteration of the AdaBoost algorithm further includes:
(iii) updating weight vectors w i s k for the training instances i of the source training sets D S k ,k=1, . . . , N based on the error for the target domain training set D T of the candidate base classifier selected as the base classifier h r (x).

7. The labeling system of claim 1 wherein the Adaboost algorithm maintains a queue PQ of candidate base classifiers across iterations of the Adaboost algorithm, and the selecting operation (ii) includes:
(ii)(a) selecting h r (x) as the candidate base classifier in the queue PQ with lowest error for the target domain training set D T ; and
(ii)(b) removing the selected candidate base classifier from the queue PQ.

8. The labeling system of claim 1 wherein the expanding of the target domain training set D T comprises:
adding to the target domain training set D T additional synthesized instances with different labels wherein the synthesized instances have initialized weight vectors for the Adaboost algorithm computed based on label probabilities generated using the prior estimate of the labels distribution.

9. A labeling method for labeling an image to be labeled belonging to a target domain, the image labeling method comprising:
computing feature vectors representing target domain training images belonging to the target domain;
labeling the target domain training images using labels selected from a set of labels to generate a target domain training set D T comprising labeled feature vectors representing the target domain training images;
receiving a plurality of source domain training sets D S 1 , . . . , D S N where N≧1 comprising feature vectors representing images belonging to source domains different from the target domain that are labeled using labels selected from the set of labels;
performing unsupervised source-target domain alignment to align the target domain training set D T and the source training sets D S k , k=1, . . . , N;
training a boost classifier f T (x)=Σ r=1 M β r h r (x) to classify an image belonging to the target domain and represented by a feature vector x, the training using the aligned target domain training set D T and plurality of source domain training sets D S 1 , . . . , D S N , the training comprising expanding the target domain training set D T based on a prior estimate of the labels distribution for the target domain and, after expanding the target domain training set D T , applying an adaptive boosting (AdaBoost) algorithm to generate the base classifiers h r (x) and the base classifier weights β r of the boost classifier f T (x), where r=1, . . . , M; and
computing a feature vector x in representing the image to be labeled; and
generating a label for the image to be labeled by operations including evaluating f T (x in ) r=1 M β r h r (x in );
wherein the feature vector computing operations, the training operation, and the generating operation are performed by an electronic data processing device.

10. The labeling method of claim 9 wherein N≧2 and the r th iteration of the AdaBoost algorithm includes (i) training candidate base classifiers h r k (x) wherein h r k (x) is trained on a training set combining the target domain training set D T and the source training set D S k and (ii) selecting h r (x) from a pool of trained candidate base classifiers based on a target domain error metric.

11. The labeling method of claim 10 wherein the operation (ii) selects h r (x) from one of:
the pool of trained candidate base classifiers h r k (x) trained during the r th iteration; and
the pool of trained candidate base classifiers trained during the r th iteration and any earlier iterations, wherein the operation (ii) further includes removing the candidate base classifier selected as h r (x) from the pool.

12. The labeling method of claim 10 wherein the operation (i) includes performing k=1, . . . , N sub-iterations in which the k th sub-iteration trains a candidate base classifier h r k (x) on a training set combining the target domain training set D T and the source training set D S k .

13. The labeling method of claim 9 wherein the set of labels is a set of object labels, and the image to be labeled, the target domain training images, and the images belonging to source domain

Assignees

• Xerox Corp

Inventors

• Chidlovskii Boris
• Csurka Gabriela

Classifications

• G06V10/774 Primary

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

• G06F18/2155

  characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06F18/214

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

• G06F18/21

  Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation · CPC title