Learning device, classification device, learning method, classification method, learning program, and classification program

A learning method executed by a computer, the method comprising:

• generating a model for classifying one or more objects present in an image, the model including
• feature extraction, wherein the feature extraction includes a plurality of convolution layers and output at least one feature map of the image, and
• coordinate estimation, wherein the coordinate estimation is configured to output coordinates of at least one region, in the image, where at least one of the one or more objects is present;
• and
• storing the model in a computer-readable medium, wherein
• generating the model includes
• acquiring a plurality of first images, from a computer-readable medium, one or more objects being present in each of the plurality of first images,
• first training feature extraction based on the plurality of acquired first images,
• causing the feature extraction trained based on the plurality of acquired first images to output at least one feature map of each of one or more second images among the plurality of first images,
• inputting, for each of the one or more second images, the at least one output feature map to the coordinate estimation and causing the coordinate estimation to output coordinates of at least one region, in the second image, where at least one of the one or more objects is present,
• extracting, for each of the one or more second images, one or more third images from the second image, the one or more third images corresponding to the at least one region indicated by the output coordinates, and
• second training the feature extraction trained based on the one or more extracted third images.

A non-transitory computer readable medium storing thereon a learning program, which when executed by a computer, causes the computer to:

• generate a model for classifying one or more objects present in an image, the model including
• feature extraction, wherein the feature extraction includes a plurality of convolution layers and output at least one feature map of the image, and
• coordinate estimation, wherein the coordinate estimation is configured to output coordinates of at least one region, in the image, where at least one of the one or more objects is present;
• and
• store the model in a computer-readable medium, wherein
• generating the model includes
• acquiring a plurality of first images, from a computer-readable medium, one or more objects being present in each of the plurality of first images,
• first training feature extraction based on the plurality of acquired first images,
• causing the feature extraction trained based on the plurality of acquired first images to output at least one feature map of each of one or more second images among the plurality of first images,
• inputting, for each of the one or more second images, the at least one output feature map to the coordinate estimation and causing the coordinate estimation to output coordinates of at least one region, in the second image, where at least one of the one or more objects is present,
• extracting, for each of the one or more second images, one or more third images from the second image, the one or more third images corresponding to the at least one region indicated by the output coordinates, and
• second training the feature extraction trained based on the one or more extracted third images.