Samsung Electronics Co., Ltd.

Apple Inc.

Primesense Ltd

Au10tix Limited

Shai Litvak

Tomer Yanir

Leonid Brailovsky

Eran Guendelman

Alexander Shpunt

Erez Sali

Micha Galor

Ofir Or

German Voronov

Yoel Yaffe

Ron Atzmon

Aviad Maizels

Amit Eisenberg

Ariel Orfaig

Ishay Goldin

Gershi Koltun

Evgeny Soloveichik

Jakob Avitan

Tomer Avidor

Yechiel Lamash

Oded Ilan-Lipowsky

Gabriel Kozakov

Nir Zoran

Ronen Gal

Eyal Hochberg

Systems and methods for structured-pruning, zero-skipping and accelerated processing of an artificial neural network (ANN) are described. The ANN may include one or more convolution layers. 2D channels in filters of the convolution layers comprise fully pruned channels (FPCs), each containing only zero weights, and mixed channels (MCs), each containing at least one non-zero weight. At least a portion of the MCs satisfy a limited zero sequence (LZS) condition limiting the number and location of zeroes in the MC. The LZS condition may be based on a number of weights that a zero-skipping circuit of a computing system for processing the ANN is configured to evaluate and skip in a single cycle. Thus, when processing the structurally-pruned ANN using the zero-skipping method, the computing system may avoid processing zero weights. This may allow speeding up the ANN processing and reducing the power required for processing.

Neural network hardware accelerator system with zero-skipping and hierarchical structured pruning methods

A data analytics system/method operative in conjunction with a data repository storing data regarding each of a multiplicity of frames including images of ID documents, including receiving at least one image generated by an image capturing device such as a camera or scanner; providing document data, derived by a hardware processor from the image capturing device, which characterizes a document depicted in the image; providing person data, derived by a hardware processor from the image, which characterizes a person who may bear the document depicted in the image; and generating inputs for identification of potential fraudulent attempts including analyzing whether the document data exists within the data regarding each of the multiplicity of images of ID documents; and/or analyzing whether the person data exists within the data regarding each of the multiplicity of images of ID documents.

System, method and computer program product for mitigating customer onboarding risk

A method for processing a digital content includes acquiring content data using a sensor. Compressed reference data is generated from the acquired content data. A hash of the compressed reference data is generated using a hashing function. The generated hash is signed using an encryption function. The acquired content data is transmitted along with the compressed reference data and the signed hash.

Robust selective image, video, and audio content authentication

A digital circuit for accelerating computations of an artificial neural network model includes a pairs selection unit that selects different subsets of pairs of input vector values and corresponding weight vector values to be processed simultaneously at each time step; a sorting unit that simultaneously processes a vector of input-weight pairs wherein pair values whose estimated product is small are routed with a high probability to small multipliers, and pair values whose estimated product is greater are routed with a high probability to large multipliers that support larger input and output values; and a core unit that includes a plurality of multiplier units and a plurality of adder units that accumulate output results of the plurality of multiplier units into one or more output values that are stored back into the memory, where the plurality of multiplier units include the small multipliers and the large multipliers.

Neural network hardware acceleration with stochastic adaptive resource allocation

A method of processing a layer in a convolutional neural network includes reading a plurality of different subsets of pairs of input vector values and corresponding weight vector values from an on-chip memory, storing the pair values in registers of a plurality of multiplication units, simultaneously processing the different subsets of pairs of values in the registers by multiplying the pairs of values and accumulating the results of the multiplications to obtain an output vector for a current layer. The input vector and the weight vector have floating point values and multiplying the pairs of values comprises adding corresponding pairs of integers in a log domain, and an input value or weight value is released from the on-chip memory after said input value or weight value is no longer needed for calculating non-computed output vector values.

Low-power hardware acceleration method and system for convolution neural network computation

Operating a computerized system includes presenting user interface elements on a display screen. A first gesture made in a three-dimensional space by a by a distal portion of an upper extremity of a user is detected while a segment of the distal portion thereof rests on a surface. In response to the first gesture, an area of the display screen selected by the user is identified, and a corresponding user interface element is displayed. After displaying the corresponding user interface element, a second gesture made by the distal portion is detected while the segment continues to rest on the surface so as to select one of the user interface elements that appears in the selected area.

Enhanced 3D interfacing for remote devices

In a method of extracting a logo mask from a video stream having a logo, a video stream including a plurality of frames is received. A plurality of online-weighted-average frames is sequentially generated using the video stream. A logo-boundary-score map is generated using a current frame among the plurality of frames and a current online-weighted-average frame among the plurality of online-weighted-average frames. The logo-boundary-score map includes a plurality of logo-boundary-score entries, each of which corresponds to one of a plurality of pixels of the video stream and includes a probability that a corresponding one of the plurality of pixels is a logo-boundary pixel. The logo mask includes a plurality of entries, each of which corresponds to one of the plurality of pixels and represents whether a corresponding one of the plurality of pixels is a logo pixel or a non-logo pixel.

Method of extracting logo mask from video stream having logo

A method for image processing includes receiving a depth image of a scene containing a human subject and receiving a color image of the scene containing the human subject. A part of a body of the subject is identified in at least one of the images. A quality of both the depth image and the color image is evaluated, and responsively to the quality, one of the images is selected to be dominant in processing of the part of the body in the images. The identified part is localized in the dominant one of the images, while using supporting data from the other one of the images.

Combined color image and depth processing

A method for processing data includes receiving a depth map of a scene containing a human hand, the depth map consisting of a matrix of pixels having respective pixel depth values. The method continues by extracting from the depth map information based on the depth values in a plurality of positions distributed over the human hand and processing the information in order to estimate respective candidate positions of the finger joints. The pose of the human hand is estimated by choosing a combination of the positions of the finger joints, responsively to anatomical constraints of the hand, that gives a hand configuration that is most anatomically probable among the candidate positions.

Learning based estimation of hand and finger pose

A method for recognizing small-font sized text including receiving digital media of a natural scene, the digital media having at least one frame that includes the small-font sized text; generating input maps having values that reflect local properties of corresponding regions in the at least one frame; and detecting regions of the at least one frame that contain the small-font sized text by integrating information from the input maps. The integrated information may include information located between border lines having active pixels therebetween and gaps having a high ratio of non-ink pixels located below a bottom border line and above a top border line in relation to a dominant direction of the text. The active pixels may be pixels having dense changes in character stroke directions.

Detecting and processing small text in digital media

High quality upscaling and denoising are required in mobile imaging devices that do not contain high quality lenses. Such is also required in order to scale up standard-definition video content for display in high-definition television screens. The disclosed method uses contextual information obtained during upscaling and/or denoising of frames. Relevant correspondences between patches within a frame and between frames, are detected, managed and exploited. The correspondence information is simultaneously used and updated while video frames are being processed. Two approaches may be used: 1. keeping, searching for and updating a database of useful patches, by adding frequently visible similar patches, aggregating high-frequency, low-noise information associated with the similar patches, and removing less-observed patches; 2. Using the high-resolution and noise-reduced information that was collected from earlier video frames, and is expressed by the output of latest processed frame, for upscaling and/or noise-reducing the next processed frame.

Image processor for and method of upscaling and denoising using contextual video information

A method for processing data includes receiving a depth map of a scene containing a human hand, the depth map consisting of a matrix of pixels having respective pixel depth values. The method continues by extracting from the depth map respective descriptors based on the depth values in a plurality of patches distributed in respective positions over the human hand, and matching the extracted descriptors to previously-stored descriptors in a database. A pose of the human hand is estimated based on stored information associated with the matched descriptors.

Learning-based estimation of hand and finger pose

A method, including presenting, by a computer system executing a non-tactile three dimensional user interface, a virtual keyboard on a display, the virtual keyboard including multiple virtual keys, and capturing a sequence of depth maps over time of a body part of a human subject. On the display, a cursor is presented at positions indicated by the body part in the captured sequence of depth maps, and one of the multiple virtual keys is selected in response to an interruption of a motion of the presented cursor in proximity to the one of the multiple virtual keys.

Virtual keyboard for a non-tactile three dimensional user interface

A method for processing data includes receiving a depth map of a scene containing a humanoid form. Respective descriptors are extracted from the depth map based on the depth values in a plurality of patches distributed in respective positions over the humanoid form. The extracted descriptors are matched to previously-stored descriptors in a database. A pose of the humanoid form is estimated based on stored information associated with the matched descriptors.

Growing community of inventors

Beit Hashmonai, Israel

Shai Litvak