Superfish Ltd.

Michael Chertok

Adi Pinhas

Alexander Lorbert

Yossi Keller

Lev Koushnir

A method for producing a graph representation of an input image, the method including the procedures of applying convolutional layers of a trained convolutional neural network on the input image, defining a receptive field of a last convolutional layer of the trained convolutional neural network as a vertex of the graph representation, defining a vector of a three dimensional output matrix of the last convolutional layer that is mapped to the receptive field as a descriptor for the vertex and determining an edge between a pair of vertices of the graph representation by applying an operator on a pair of descriptors respective of the pair of vertices.

Graph image representation from convolutional neural networks

A method for representing an input image includes the steps of applying a trained neural network on the input image, selecting a plurality of feature maps, determining a location of each of the plurality of feature maps in an image space of the input image, defining a plurality of interest points of the input image, and employing the plurality of interest points for representing the input image for performing a visual task. The plurality of feature maps are selected of an output of at least one selected layer of the trained neural network according to values attributed to the plurality of feature maps by the trained neural network. The plurality of interest points of the input image are defined based on the locations corresponding to the plurality of feature maps.

Neural network image representation

A system and method for processing a digital image. A digital image and a definition of a segment therein may be obtained. A sampling area may be defined wherein the sampling area at least partly overlaps with the segment. A characteristic value for the sampling area may be determined The image may be represented based on the characteristic value. Other embodiments are described and claimed.

System and method for processing a digital image

A method for determining a matching score between a first set of nfeature points, and a second set of nfeature points includes producing a triple-wise affinity tensor, determining a leading eigenvector of the triple-wise affinity tensor, iteratively producing a binary optimal assignment vector and determining a matching score between the first set of feature points and the second set of feature points. The triple-wise affinity tensor includes the affinity score of assignments of triplets of feature points of the first set of feature points and triplets of feature points of the second set of feature points. The procedure of iteratively producing a binary optimal assignment vector is performed by discretization of the leading eigenvector. The procedure of determining a matching score is performed according to the triple-wise affinity tensor and according to the optimal assignment vector.

Method for organizing a database of images and retrieving images from that database according to a query image

A system for generating virtual interest points in an image, according to detected interest points in that image is detected by an interest point detector. The system includes: an interest point density map producer and an interest point generator, the interest point density map producer receiving the image including the detected interest points, the interest point density map producer extracts a density map of the detected interest points within the image. The interest point generator is coupled with the interest point density map producer, the interest point generator receiving the image including the detected interest points and the interest point density map. The interest point generator determines image areas for generating virtual interest points, the interest point generator generating at least one virtual interest point at the determined areas according to at least the position of one of the detected interest points, the virtual interest point not being detected by the interest point detector.

Increasing interest point coverage in an image

A system and method for identifying a region of interest in a digital image. A first and second images of a scene may be obtained from a respective first and second points of view. Following an acquisition of a first image from a first point of view, a subsequent image may be automatically acquired upon determining that a second view point is achieved. Based on two or more images of a scene, a background object may be removed from an image to produce an image that only includes a foreground object or a region of interest.

System and method for identifying a region of interest in a digital image

A method for generating a multi scale representation of an input image, the method comprising the procedures of: estimating a scale factor corresponding to said input image; determining a set of Gaussian difference kernels according to said estimated scale factor, and according to a predetermined set of Gaussian kernels; and generating a multi-scale representation of said input image by applying each of said set of Gaussian difference kernels on said input image.

Multi-scale representation of an out of focus image

System for generating an edge neighborhood descriptor for describing the surrounding of an interest point according to the closest edges includes a sector determiner, a closest edge determiner, and an edge neighborhood descriptor constructor, the closest edge determiner is coupled between the sector determiner and the edge neighborhood descriptor constructor. The sector determiner determines N sectors, surrounding the interest point. The closest edge determiner determines for each of the N sectors, the edge pixel closest to the interest point, according to at least one binary edge map. The edge neighborhood descriptor constructor constructs the edge neighborhood descriptor such that the length of the radius of each of the N sectors is determined according to at least the distance from the interest point to the edge pixel closest to the interest point within the sector, the edge neighborhood descriptor includes the N sectors.

Edge neighborhood descriptor

Method for organizing a set of images into subsets of images, the method including the following procedures producing a respective model according to a plurality of feature points and the geometric relations between the feature points for each of the images, determining a similarity index between each pair of the images, according to the respective model of each image of the pair of the images, producing a distance matrix, according to the similarity index between each pair of the images, producing a set of coordinates, according to the distance matrix and sorting the images in plurality of dimensions, according to the set of coordinates.

Method and a system for organizing an image database

System for providing a mobile user, object related information related to an object visible thereto, the system including a camera directable toward the object, a local interest points and semi global geometry (LIPSGG) extraction processor, and a remote LIPSGG identifier, the camera acquiring an image of at least a portion of the object, the LIPSGG extraction processor being coupled with the camera, the LIPSGG extraction processor extracting an LIPSGG model of the object from the image, remote LIPSGG identifier being coupled with the LIPSGG extraction processor via a network, the remote LIPSGG identifier receiving the LIPSGG model from the LIPSGG extraction processor, via the network, the remote LIPSGG identifier identifying the object according to the LIPSGG model, the remote LIPSGG identifier retrieving the object related information, the remote LIPSGG identifier providing the object related information to the mobile user operating the camera.

Growing community of inventors

Petah Tikva, Israel

Michael Chertok