Classification device, classification method, program, and information recording medium

A method of classification comprising the steps, performed by a classification device, of:

• acquiring a plurality of learning images Si,1, Si,2, . . . , which are obtained by dividing each learning photograph Si of a plurality of learning photographs S1, S2, . . . , associating the plurality of learning images with a group Gz(i) to which a learning target captured on each learning photograph Si belongs among a plurality of groups G1, G2, . . . , GL;
• learning a first model for calculating a plurality of compressed vectors
• v1,1, v1,2, . . . ,
• v2,1, v2,2, . . . ,
• . . . ,
• obtained by dimensionality reduction from the plurality of acquired learning images
• S1,1, S1,2. . . ,
• S2,1, S2,2. . . ,
• . . . ,
• a class Cy(i,j) among a plurality of classes C1, C2, . . . . , CM, wherein each learning image Si,j of the plurality of learning images and each compressed vector vi,j of the plurality of compressed vectors are classified into the class Cy(i,j), and a plurality of M-dimensional score vectors ui,j which represent a nearness of each compressed vector vi,j to each of the plurality of classes;
• calculating an M-dimensional representative degree vector rh that represents the degree to which each of the plurality of classes C1, C2, . . . . , CM represents each group Gh of the plurality of groups G1, G2, . . . , GL from one or more score vectors ui,j|z(i)=h calculated for one or more compressed vectors vi,j|z(i)=h associated with each group Gh, and associating each class Ck of the plurality of classes with a group Gx(k) from among the plurality of groups G1, G2, . . . , GL and another group G0 based on the calculated representative degree vectors r1, r2, rL;
• receiving a plurality of determination images obtained by dividing a determination photograph on which a determination target is captured;
• and
• classifying each of the received plurality of determination images to one of the plurality of classes, using the learned first model, and identifying the group associated with the class into which the image is classified.

A program causing a computer to function as:

• an acquirer that acquires a plurality of learning images Si,1, Si,2, . . . , which are obtained by dividing each learning photograph Si of a plurality of learning photographs S1, S2, . . . , the acquirer associating the plurality of learning images with a group Gz(i) to which a learning target captured on each learning photograph Si belongs among a plurality of groups G1, G2, . . . , GL;
• a first learner that learns a first model for calculating a plurality of compressed vectors
• v1,1, v1,2, . . . ,
• v2,1, v2,2, . . . ,
• . . . ,
• obtained by dimensionality reduction from the plurality of acquired learning images
• S1,1, S1,2, . . . ,
• S2,1, S2,2, . . . ,
• . . . ,
• a class Cy(i,j) among a plurality of classes C1, C2, . . . . , CM, wherein each learning image Si,j of the plurality of learning images and each compressed vector vi,j of the plurality of compressed vectors are classified into the class Cy(i,j), and a plurality of M-dimensional score vectors ui,j which represent a nearness of each compressed vector vi,j to each of the plurality of classes;
• an associator that calculates an M-dimensional representative degree vector rh that represents the degree to which each of the plurality of classes C1, C2, . . . . , CM represents each group Gh of the plurality of groups G1, G2, . . . , GL from one or more score vectors ui,j|z(i)=h calculated for one or more compressed vectors vi,j|z(i)=h associated with each group Gh, and associates each class Ck of the plurality of classes with a group Gx(k) from among the plurality of groups G1, G2, . . . , GL and another group G0 based on the calculated representative degree vectors r1, r2, . . . , rL;
• a receiver that receives a plurality of determination images obtained by dividing a determination photograph on which a determination target is captured;
• and
• an identifier that classifies each of the received plurality of determination images to one of the plurality of classes, using the learned first model, and identifies the group associated with the class into which the image is classified.