Image recognition apparatus, image recognition method, and a learning data set generation apparatus
In an image recognition apparatus, a processor performs, based on an input image and using an image recognition model, a plurality of object detection processes to detect as an object detection region a region in the input image where a recognition target object is judged to be present. In the plurality of object detection processes, a plurality of mutually different image recognition models are used. The processor generates inference result data according to the degree of overlap among a plurality of object detection regions detected in the plurality of object detection processes.
1. An image recognition apparatus comprising a processor configured to make an inference to recognize a recognition target object in an input image, wherein
the processor is configured to
perform, based on the input image and using a plurality of mutually different image recognition models, a plurality of object detection processes to detect as an object detection region a region in the input image where the recognition target object is judged to be present, and
generate inference result data indicating a result of the inference in accordance with a degree of overlap among a plurality of object detection regions detected in the plurality of object detection processes,
the processor is configured to generate, in each object detection process, object detection data including
class data indicating a kind of an object in the object detection region,
position/shape data of the object detection region, and
a reliability of the result of the object detection process,
each object detection process is a single-class object detection process dealing with a single kind of object as the recognition target object, the recognition target objects of the plurality of object detection processes being of a same kind, and
the processor is configured to
identify a maximum reliability among a plurality of reliabilities derived in the plurality of object detection processes and
set to the inference result data the class data and the position/shape data corresponding to the maximum reliability if the degree of overlap is equal to or higher than a predetermined overlap threshold value.
2. The image recognition apparatus according to claim 1 , wherein
the processor is configured to set to the inference result data the class data and the position/shape data of the object detection data corresponding to the maximum reliability
if the degree of overlap is equal to or higher than the overlap threshold value or
if the degree of overlap is lower than the overlap threshold value and in addition the maximum reliability is equal to or higher than a predetermined reference reliability.
3. The image recognition apparatus according to claim 2 , wherein
the processor is configured to set particular data as the inference result data if the degree of overlap is lower than the overlap threshold value and in addition the maximum reliability is lower than the reference reliability.
4. The image recognition apparatus according to claim 1 , wherein
the plurality of image recognition models employ mutually different algorithms for detecting the recognition target object, or
each image recognition model is constructed as a neural network, the plurality of image recognition models being constructed as neural networks having mutually different configurations, or
the plurality of image recognition models are constructed through machine learning in mutually different environments.
5. A learning data set generation apparatus comprising the image recognition apparatus according to claim 1 , wherein
the learning data set generation apparatus is configured to generate a learning data set using a plurality of input images and the inference result data with respect to each input image.
6. An image recognition apparatus comprising a processor configured to make an inference to recognize a recognition target object in an input image, wherein
the processor is configured to
perform, based on the input image and using a plurality of mutually different image recognition models, a plurality of object detection processes to detect as an object detection region a region in the input image where the recognition target object is judged to be present, and
generate inference result data indicating a result of the inference in accordance with a degree of overlap among a plurality of object detection regions detected in the plurality of object detection processes,
the processor is configured to generate, in each object detection process, object detection data including
class data indicating a kind of an object in the object detection region,
position/shape data of the object detection region, and
a reliability of the result of the object detection process,
each object detection process is a multi-class object detection process dealing with a plurality of kinds of object as the recognition target object, and
the processor is configured to
identify a maximum reliability among a plurality of reliabilities derived in the plurality of object detection processes and
set to the inference result data the class data and the position/shape data of the object detection data corresponding to the maximum reliability if the degree of overlap is equal to or higher than a predetermined overlap threshold value in a case where a plurality of regions where through the plurality of object detection processes a same kind of object is judged to be present are the plurality of object detection regions.
7. The image recognition apparatus according to claim 6 , wherein
the processor is configured to set to the inference result data the class data and the position/shape data of the object detection data corresponding to the maximum reliability
if the degree of overlap is equal to or higher than the overlap threshold value or
if the degree of overlap is lower than the overlap threshold value and in addition the maximum reliability is equal to or higher than a predetermined reference reliability
in the case where the plurality of regions where through the plurality of object detection processes the same kind of object is judged to be present are the plurality of object detection regions.
8. The image recognition apparatus according to claim 7 , wherein
the processor is configured to set particular data as the inference result data if the degree of overlap is lower than the overlap threshold value and in addition the maximum reliability is lower than the reference reliability in the case where the plurality of regions where through the plurality of object detection processes the same kind of object is judged to be present are the plurality of object detection regions.
9. The image recognition apparatus according to claim 6 , wherein
the plurality of image recognition models employ mutually different algorithms for detecting the recognition target object, or
each image recognition model is constructed as a neural network, the plurality of image recognition models being constructed as neural networks having mutually different configurations, or
the plurality of image recognition models are constructed through machine learning in mutually different environments.
10. A learning data set generation apparatus comprising the image recognition apparatus according to claim 6 , wherein
the learning data set generation apparatus is configured to generate a learning data set using a plurality of input images and the inference result data with respect to each input image.
11. An image recognition method involving making an inference to recognize a recognition target object in an input image, the method comprising:
performing, based on the input image and using a plurality of mutually different image recognition models, a plurality of object detection processes to detect as an object detection region a region in the input image where the recognition target object is judged to be present; and
generating inference result data indicating a result of the inference in accordance with a degree of overlap among a plurality of object detection regions detected in the plurality of object detection processes,
generating, in each object detection process, object detection data including
class data indicating a kind of an object in the object detection region,
position/shape data of the object detection region, and
a reliability of the result of the object detection process,
each object detection process being a single-class object detection process dealing with a single kind of object as the recognition target object, the recognition target objects of the plurality of object detection processes being of a same kind,
identifying a maximum reliability among a plurality of reliabilities derived in the plurality of object detection processes and
setting to the inference result data the class data and the position/shape data of the object detection data corresponding to the maximum reliability if the degree of overlap is equal to or higher than a predetermined overlap threshold value.
12. An image recognition method involving making an inference to recognize a recognition target object in an input image, the method comprising:
performing, based on the input image and using a plurality of mutually different image recognition models, a plurality of object detection processes to detect as an object detection region a region in the input image where the recognition target object is judged to be present; and
generating inference result data indicating a result of the inference in accordance with a degree of overlap among a plurality of object detection regions detected in the plurality of object detection processes,
generating, in each object detection process, object detection data including
class data indicating a kind of an object in the object detection region,
position/shape data of the object detection region, and
a reliability of the result of the object detection process,
each object detection process being a multi-class object detection process dealing with a plurality of kinds of object as the recognition target object, and
identifying a maximum reliability among a plurality of reliabilities derived in the plurality of object detection processes and
setting to the inference result data the class data and the position/shape data of the object detection data corresponding to the maximum reliability if the degree of overlap is equal to or higher than a predetermined overlap threshold value in a case where a plurality of regions where through the plurality of object detection processes a same kind of object is judged to be present are the plurality of object detection regions.