公开(公告)号：US12094187B2

公开(公告)日：2024-09-17

申请号：US17619208

申请日：2019-06-17

申请人： NIPPON TELEGRAPH AND TELEPHONE CORPORATION

发明人： Kaori Kumagai ,  Jun Shimamura ,  Atsushi Sagata

IPC分类号： G06V10/762 ,  G06V10/22 ,  G06V10/764 ,  G06V10/77 ,  G06V10/82

CPC分类号： G06V10/7715 ,  G06V10/22 ,  G06V10/762 ,  G06V10/764 ,  G06V10/82

摘要： An identification result explanation device calculates, for each of feature maps output individually from a plurality of filters used in specified layers of a CNN, a weight representing a degree of association with a result of identification for an input image by the CNN. The identification result explanation device outputs transposed feature maps obtained by transposing, for each of clusters, the feature maps in the cluster based on a result of classification of each of the feature maps for the input image classified into any of the clusters and on the weight calculated for each of the feature maps. The identification result explanation device uses the transposed feature maps in each of the clusters to retrieve, in a storage unit, each of the clusters including the feature maps linked to the same filters as those linked to the transposed feature maps for the input image and selects.

公开(公告)号：US11416710B2

公开(公告)日：2022-08-16

申请号：US16971635

申请日：2019-02-25

申请人： NIPPON TELEGRAPH AND TELEPHONE CORPORATION

发明人： Kaori Kumagai ,  Yukito Watanabe ,  Jun Shimamura ,  Tetsuya Kinebuchi

IPC分类号： G06K9/62 ,  G06N3/04 ,  G06N3/08 ,  G06V10/22

摘要： The present invention relates to representing image features used by a convolutional neural network (CNN) to identify concepts in an input image. The CNN includes a plurality of filters in each of a plurality of layers. The method generates the CNN based on a set of images for training with predetermined concepts in regions of the set of images. For a select layer of the CNN, the method generates integrated maps, Each integrated map is based on a set of feature maps in a cluster and relevance between the set of feature maps for the select layer and a region representing one of the features in the image data. The method provides a pair of a feature representation visualization image of a feature in the select layer and a concept information associated with the integration map.