公开(公告)号：GB2547752A

公开(公告)日：2017-08-30

申请号：GB201621726

申请日：2016-12-20

Applicant: IBM

Inventor： SHARATHCHANDRA UMAPATHIRAO PANKANTI ,  ARVIND KUMAR ,  JANUSZ MARECKI ,  BAN KAWAS

IPC: G06T7/70

Abstract: A received image (1301) is divided into patches, which may have different sizes. A cluster-direction sequence (see fig. 6) is generated (1302) for each of a plurality of saccadic paths along the patches, the paths being given by a policy matrix (see figs. 8 & 9). The sequences are used to identify an object in the image. Path sequence generation using the matrix may comprise assigning (1303) a likelihood (which may be weighted (1304) using a total frequency of occurrence of the sequence in the matrix for a given class) that the image belongs to each class defined in the matrix, and identifying (1305) the object using a likelihood average over the sequences. Image context, e.g. previous observations or goal, may be used. The patches may be clustered into groups, e.g. using k-means clustering algorithm (see figs. 4&5). Specific salient features in an image may thus be identified, leading to a classification of the image by judging whether the salient features can identify a unique class of object (e.g. 9 MNIST digits, or image classes in ImageNet, e.g. cats, planes etc.) in the image with high probability. Classification may occur through progressive exclusion of other classes.

公开(公告)号：GB2547752B

公开(公告)日：2018-01-24

申请号：GB201621726

申请日：2016-12-20

Applicant: IBM

Inventor： SHARATHCHANDRA UMAPATHIRAO PANKANTI ,  ARVIND KUMAR ,  JANUSZ MARECKI ,  BAN KAWAS

IPC: G06T7/70

Abstract: A method of operating an image detection device includes receiving an image, dividing the image into a plurality of patches, grouping ones of the plurality of patches, generating a set of saccadic paths through the plurality of patches of the image, generating a cluster-direction sequence for each saccadic path, generating a policy function for identifying an object in a new image using a combination of the cluster-direction sequences, and operating the image detection device using the policy function to identify an object in the new image.