Inference device, inference method, and a non-transitory computer-readable medium
An image signal indicating an inference target image in which a detection target appears is acquired when a domain of the inference target image is different from a domain of a training image or a recognition task of the inference target image is different from a pre-learned task. The image signal is provided to a trained learning model, and from the learning model, an inference time feature amount obtained by combining feature amounts of the detection target in the inference target image after the feature amounts is blurred is acquired. The detection target in the inference target image is recognized on the basis of a representative feature amount that is a registered feature amount of the detection target in an image for conversion in which a domain and a recognition task of the image are the same as those of the inference target image, and the inference time feature amount.
1 . An inference device comprising: processing circuitry
an image signal acquiring unit to acquire an image signal indicating an inference target image that is an image in which a detection target object appears in a case where one or more of a condition that a domain of the inference target image is different from a domain of a training image and a condition that a recognition task of the inference target image is different from a pre-learned task are satisfied,
to provide the image signal to a learning model in which learning of the training image has been completed, and acquire, from the learning model, an inference time feature amount that is obtained by combining a plurality of feature amounts of the detection target object appearing in the inference target image after each of the plurality of feature amounts is blurred, and
to recognize the detection target object appearing in the inference target image on a basis of a representative feature amount that is a registered feature amount of the detection target object appearing in an image for conversion in which each of a domain and a recognition task of the image is the same as that of the inference target image, and the inference time feature amount.
2 . The inference device according to claim 1 , wherein the processing circuitry acquires an image signal indicating the image for conversion,
provides the image signal indicating the image for conversion to the learning model, and acquires, from the learning model, a representative feature amount that is obtained by combining a plurality of feature amounts of the detection target object appearing in the image for conversion after each of the plurality of feature amounts is blurred, and
registers a representative feature amount.
3 . The inference device according to claim 1 , wherein the processing circuitry calculates similarity between a feature vector indicating the representative feature amount and a feature vector indicating an inference time feature amount, and recognizes a detection target object appearing in the inference target image on a basis of the similarity.
4 . The inference device according to claim 1 , wherein the processing circuitry compares representative feature amounts of a plurality of detection target objects having different types from each other with the inference time feature amount, performs specification of a representative feature amount corresponding to the inference time feature amount among the representative feature amounts of the plurality of detection target objects, and recognizes a type of the detection target object as recognition of the detection target object appearing in the inference target image on a basis of a result of the specification.
5 . The inference device according to claim 1 , wherein the processing circuitry compares representative feature amounts of a plurality of detection target objects having different presence regions from each other with the inference time feature amount, performs specification of a representative feature amount corresponding to the inference time feature amount among the representative feature amounts of the plurality of detection target objects, and recognizes a region where the detection target object is present as recognition of the detection target object appearing in the inference target image on a basis of a result of the specification of the representative feature amount.
6 . The inference device according to claim 1 , wherein the processing circuitry compares representative feature amounts of a plurality of detection target objects having types and presence regions different from each other with the inference time feature amount, performs specification of a representative feature amount corresponding to the inference time feature amount among the representative feature amounts of the plurality of detection target objects, and recognizes each of a type and a presence region of a detection target object as recognition of the detection target object appearing in the inference target image on a basis of a result of the specification of the representative feature amount.
7 . The inference device according to claim 1 , wherein the learning model includes deep neural networks (DNNs), and
the processing circuitry provides the image signal to the DNNs and acquires the inference time feature amount from the DNNs.
8 . An inference method comprising:
acquiring an image signal indicating an inference target image that is an image in which a detection target object appears in a case where one or more of a condition that a domain of the inference target image is different from a domain of a training image and a condition that a recognition task of the inference target image is different from a pre-learned task are satisfied;
providing the image signal to a learning model in which learning of the training image has been completed, and acquiring, from the learning model, an inference time feature amount that is obtained by combining a plurality of feature amounts of the detection target object appearing in the inference target image after each of the plurality of feature amounts is blurred; and
recognizing the detection target object appearing in the inference target image on a basis of a representative feature amount that is a registered feature amount of the detection target object appearing in an image for conversion in which each of a domain and a recognition task of the image is the same as that of the inference target image, and the inference time feature amount.
9 . A non-transitory computer-readable medium storing a program including instructions that, when executed by a processor, causes a computer to execute a process, the process including:
to acquire an image signal indicating an inference target image that is an image in which a detection target object appears in a case where one or more of a condition that a domain of the inference target image is different from a domain of a training image and a condition that a recognition task of the inference target image is different from a pre-learned task are satisfied;
to provide the image signal to a learning model in which learning of the training image has been completed, and acquire, from the learning model, an inference time feature amount that is obtained by combining a plurality of feature amounts of the detection target object appearing in the inference target image after each of the plurality of feature amounts is blurred; and
to recognize the detection target object appearing in the inference target image on a basis of a representative feature amount that is a registered feature amount of the detection target object appearing in an image for conversion in which each of a domain and a recognition task of the image is the same target as that of the inference target image, and the inference time feature amount.