CORRESPONDING POINT DETECTION DEVICE, AND CORRESPONDING POINT DETECTION METHOD AND PROGRAM

To detect feature points of a plurality of images even when the direction of a subject included in an image is not learnt.SOLUTION: A corresponding point detection device has: a propagation control part which propagates a plurality of processing layers included in a learning model to a first image and a second image; an extraction part which extracts one or more first image outputs based upon the first image and one or more second image outputs based upon the second image; and a feature point detection part which detects, on the basis of the one or more first image outputs and the one or more second image outputs, one or more first image feature points as feature points included in the first image and one or more second image feature points as feature points included in the second image. The learning model has: a plurality of rotation layers which rotate received input data at different angles respectively and a convolution layer which applies a convolution filter to a plurality of conversion data that the plurality of rotation layers convert.SELECTED DRAWING: Figure 5 【課題】画像に含まれる被写体の向きが未学習であっても複数の画像における特徴点を検出する。【解決手段】対応点検出装置は、第１画像と第２画像とのそれぞれに、学習モデルに含まれる複数の処理層を伝搬させる伝搬制御部と、前記第１画像に基づく一以上の第１画像出力と、前記第２画像に基づく一以上の第２画像出力とを抽出する抽出部と、前記一以上の第１画像出力と前記一以上の第２画像出力とに基づいて、前記第１画像に含まれる特徴点である一以上の第１画像特徴点と、前記第２画像に含まれる特徴点である一以上の第２画像特徴点とを検出する特徴点検出部と、を有し、前記学習モデルは、入力された入力データにそれぞれ異なる角度の回転を施す複数の回転層と、前記複数の回転層が変換した複数の変換データに畳み込みフィルタを適用する畳み込み層とを有する。【選択図】図５