Intel Corporation

Gilad Michael

Furkan Isikdogan

Chyuan-Tyng Wu

Bhavin V Nayak

Joao Peralta Moreira

Timo Gerasimow

Aleksandar Sutic

Sushma Rao

Zoran Zivkovic

Hila Barel

Avi Kalderon

Shaul Oron

German Voronov

Liron Ain-Kedem

Jun Nishimura

Chaitanya R Gandra

Edmond Chalom

Hadas Dahan

Pavel Kounitsky

Leo Isikdogan

Nir Darshan

Ben Berlin

Masayoshi Asama

An apparatus to facilitate partially-frozen neural networks for efficient computer vision systems is disclosed. The apparatus includes a frozen core to store fixed weights of a machine learning model, one or more trainable cores coupled to the frozen core, the one or more trainable cores comprising multipliers for trainable weights of the machine learning model, and wherein the alpha blending layer includes a trainable alpha blending parameter, and wherein the trainable alpha blending parameter is a function of a trainable parameter, a sigmoid function, and outputs of frozen and trainable blocks in a preceding layer of the machine learning model.

Partially-frozen neural networks for efficient computer vision systems

An example apparatus for adjusting eye gaze in images one or more processors to execute instructions to bidirectionally train a neural network; access a target angle and an input image, the input image including an eye in a first position; generate a vector field with the neural network; and generate a gaze-adjusted image based on the vector field, the gaze-adjusted image including the eye in a second position.

Generating gaze corrected images using bidirectionally trained network

An apparatus, method, and a computer readable medium for attenuating visual artifacts in processed images. An annotated dataset of images to be processed by an image processing system is created. An adversarial control network is trained to operate as an image quality expert in classifying images. After the adversarial control network has been trained, the adversarial control network is used to supervise the image processing system on-the-fly.

Attenuating visual artifacts of image processing systems using adversarial networks on-the-fly

An example apparatus for processing images includes a trainable vision scaler to receive an image. The trainable vision scaler is to generate output including a feature map or an enhanced image based on the image. The trainable vision scaler is to transmit the output to a computer vision network. The computer vision network is trained to perform a computer vision task using the output.

Trainable vision scaler

In an example method for training image signal processors, a reconstructed image is generated via an image signal processor based on a sensor image. An intermediate loss function is generated based on a comparison of an output of one or more corresponding layers of a computer vision network and a copy of the computer vision network. The output of the computer vision network is based on the reconstructed image. An image signal processor is trained based on the intermediate loss function.

Training image signal processors using intermediate loss functions

An example apparatus for processing images includes a hybrid infinite impulse response-finite impulse response (IIR-FIR) convolution block to receive an image and generate processed image information. The hybrid IIR-FIR convolution block includes a vertical infinite impulse response (IIR) component to approximate a vertical convolution when processing the image.

Processing images using hybrid infinite impulse response (TTR) and finite impulse response (FIR) convolution block

An example apparatus for correcting gaze in images includes an image receiver to receive an image comprising an eye and a target angle set to a center. The apparatus also includes a bidirectionally trained convolutional neural network (CNN) to receive the image and the target angle from the image receiver and generate a vector field and a brightness map based on the image and the target angle. The apparatus further includes an image corrector to generate a gaze corrected image based on the vector field and the brightness map.

Techniques related to computer vision tasks are discussed. Such techniques include applying a pretrained non-linear transform and pretrained details boosting factor to generate an enhanced image from an input image and reducing the bit depth of the enhanced image prior to applying a pretrained computer vision network to perform the computer vision task.

Local tone mapping to reduce bit depth of input images to high-level computer vision tasks

A mechanism is described for facilitating automatic tuning of image signal processors using reference images in image processing environments, according to one embodiment. A method of embodiments, as described herein, includes one or more processors to: receive images associated with one or more scenes captured by one or more cameras; access tuning parameters associated with functionalities within an image signal processor (ISP) pipeline; generate reference images based on the tuning parameters, wherein a reference image is associated with an image for each functionality within the ISP pipeline; and automatically tune the ISP pipeline based on selection of one or more of the reference images for one or more of the images for one or more of the functionalities.

Automatic tuning of image signal processors using reference images in image processing environments

Techniques related to generating dictionaries for example based image processing algorithms are discussed. Such techniques may include iteratively performing example based image processing for candidate look up entries of candidate pairs from a training set database using a current dictionary to determine a test result for each of the look up entries of the candidate pairs and selecting one or more of the candidate pairs for entry in a resultant dictionary based on an error between the test result and a predetermined result entry for the candidate pairs.

Dictionary generation for example based image processing

Techniques related to approximate nearest neighbor searching are discussed. Such techniques may include traversing an approximate nearest neighbor search tree from root node to a resultant leaf node while maintaining a priority queue of best matches, determining candidate entries for evaluation based on the resultant leaf node, and generating search results based on the priority queue and the candidate entries.

Hybrid nearest neighbor search tree with hashing table

Techniques related to generating downscaled image or image frame data in a luma chroma separated color space for an image or video pipeline architecture are discussed. Such techniques may include converting input image data to the luma chroma separated color space based on adaptive color coefficients determined based on an illumination indicator associated with the input image data and storing downscaled color converted image data to an input image buffer of the image or video pipeline.

Adaptive linear luma domain video pipeline architecture

A processor includes a front end including a decoder, an execution unit including a shift-sum multiplier (SSM), and a retirement unit. The decoder includes logic identify a multiplication instruction to multiply a first number and a second number. The execution unit includes logic to, based on the instruction, access a look-up table based on the second number to determine a plurality of shift parameters and one or more flag parameters. The SSM includes logic to use the shift parameters to shift the first number to determine a plurality of partial products, and the flag parameters to determine signs of the partial products. The SSM also includes logic to sum the partial products to yield a result of the multiplication instruction.

Growing community of inventors

Sunnyvale, CA, United States of America

Gilad Michael