Learning apparatus, learning method, inference apparatus, inference method, and recording medium
A learning apparatus includes a metric space learning unit and a case example storage unit. The metric space learning unit learns a metric space including feature vectors extracted from sets of attribute-attached image data for each combination of different attributes by using the sets of attribute-attached image data to which pieces of attribute information are added. The case example storage unit calculates feature vectors from sets of case example image data, and stores the feature vectors as case examples associated with the metric space.
1. A system comprising a learning apparatus and an inference apparatus for using outputs from the learning apparatus,
wherein the learning apparatus comprises:
a first memory storing first instructions; and
one or more first processors configured to execute the first instructions to:
learn, for each of a plurality of combinations of different attributes, a metric space including feature vectors which are extracted from sets of attribute-attached image data to which pieces of attribute information are respectively added, by using the sets of attribute-attached image data, such that a plurality of the metric spaces are learned for the plurality of combinations of different attributes; and
calculate, for each combination of different attributes, feature vectors from sets of case example image data and store the calculated feature vectors as case examples associated with the metric space in the first memory, and
wherein the inference apparatus comprises:
a second memory storing second instructions; and
one or more second processors configured to execute the second instructions to:
store, in the second memory and for each combination of different attributes, the feature vectors of the sets of case example image data as the case examples in association with the metric space learned for each combination of different attributes;
select one metric space by evaluating the plurality of metric spaces using a feature vector of selection image data;
recognize inference image data based on feature vectors extracted from the inference image data and the case examples associated with the one metric space; and
output a recognition result.
2. The system according to claim 1 , wherein the second processor identifies selection image data of an existing class by using each of the plurality of metric spaces and determines the one metric space having a highest degree of matching with respect to a training label for the selection image data of the existing class.
3. The system according to claim 1 , wherein the second processor determines, as the recognition result, a class of the case example closest to the feature vectors of the inference image data in the one metric space among the case examples stored in the second memory.
4. The system according to claim 3 , wherein the second processor outputs, as an inference result, a training label, additional information, and image data of the closest case example, in addition to outputting the recognition result.
5. The system according to claim 1 , wherein the second processor is further configured to perturb the inference image data, and
the second processor recognizes the inference image data using feature vectors of the perturbed inference image data.
6. The system according to claim 1 , wherein the second processor is further configured to perturb the feature vectors of the inference image data, and
the second processor recognizes the inference image data using the perturbed feature vectors.
7. A method performed by a system comprising a learning apparatus and an inference apparatus for using outputs from the learning apparatus, the method comprising:
by the learning apparatus:
learning, for each of a plurality of combinations of different attributes, a metric space including feature vectors which are extracted from sets of attribute-attached image data to which pieces of attribute information are respectively added, by using the sets of attribute-attached image data, such that a plurality of the metric spaces are learned for the plurality of combinations of different attributes; and
calculating, for each combination of different attributes, feature vectors from sets of case example image data and store the calculated feature vectors as case examples associated with the metric space in a first memory of the learning apparatus, and
by the inference apparatus:
storing, in a second memory of the inference apparatus and for each combination of different attributes, the feature vectors of the sets of case example image data as the case examples in association with the metric space learned for each combination of different attributes;
selecting one metric space by evaluating the plurality of metric spaces using a feature vector of selection image data;
recognizing inference image data based on feature vectors extracted from the inference image data and the case examples associated with the one metric space; and
outputting a recognition result.
8. A non-transitory computer-readable recording medium storing a program executable by a system to perform processing, the system comprising a learning apparatus and an inference apparatus for using outputs from the learning apparatus, the processing comprising:
by the learning apparatus:
learning, for each of a plurality of combinations of different attributes, a metric space including feature vectors which are extracted from sets of attribute-attached image data to which pieces of attribute information are respectively added, by using the sets of attribute-attached image data, such that a plurality of the metric spaces are learned for the plurality of combinations of different attributes; and
calculating, for each combination of different attributes, feature vectors from sets of case example image data and store the calculated feature vectors as case examples associated with the metric space in a first memory of the learning apparatus, and
by the inference apparatus:
storing, in a second memory of the inference apparatus and for each combination of different attributes, the feature vectors of the sets of case example image data as the case examples in association with the metric space learned for each combination of different attributes;
selecting one metric space by evaluating the plurality of metric spaces using a feature vector of selection image data;
recognizing inference image data based on feature vectors extracted from the inference image data and the case examples associated with the one metric space; and
outputting a recognition result.