Leland Stanford Junior University

Google Inc.

Yuxin Hu

Brian A Hargreaves

Bruce L Daniel

Evan G Levine

Mark Alan Duchaineau

Xiang Zhu

Methods, systems, devices, and tangible non-transitory computer readable media for haze reduction are provided. The disclosed technology can include generating feature vectors based on an input image including points. The feature vectors can correspond to feature windows associated with features of different portions of the points. Based on the feature vectors and a machine-learned model, a haze thickness map can be generated. The haze thickness map can be associated with an estimate of haze thickness at each of the points. Further, the machine-learned model can estimate haze thickness associated with the features. A refined haze thickness map can be generated based on the haze thickness map and a guided filter. A dehazed image can be generated based on application of the refined haze thickness map to the input image. Furthermore, a color corrected dehazed image can be generated based on performance of color correction operations on the dehazed image.

Spatially varying reduction of haze in images

A diffusion-weighted magnetic resonance imaging (MRI) method acquires MRI scan data from a multi-direction, multi-shot, diffusion-weighted MRI scan, and jointly reconstructs from the MRI scan data 1) magnitude images for multiple diffusion-encoding directions and 2) phase images for multiple shots and multiple diffusion-encoding directions using an iterative reconstruction method. Each iteration of the iterative reconstruction method comprises a gradient calculation, a phase update to update the phase images, and a magnitude update to update the magnitude images. Each iteration minimizes a cost function comprising a locally low-rank (LLR) regularization constraint on the magnitude images from the multiple diffusion-encoding directions.

Diffusion-weighted MRI with magnitude-based locally low-rank regularization

Multi-shot diffusion-weighted magnetic resonance imaging acquires multiple k-space segments of diffusion-weighted MRI data, estimates reconstructed multi-shot diffusion weighted images, and combines the estimated images to obtain a final reconstructed MRI image. The estimation of images iteratively calculates updated multi-shot images from the multiple k-space segments and current multi-shot images using a convex model without estimating motion-induced phase, constructs multiple locally low-rank spatial-shot matrices from the updated multi-shot images, and calculates current multi-shot images from spatial-shot matrices.

Motion robust reconstruction of multi-shot diffusion-weighted images without phase estimation via locally low-rank regularization

A method of magnetic resonance imaging performs a scan by a magnetic resonance imaging system to acquire k-space data; applies the k-space data as input to an unrolled convolutional neural network comprising multiple iterations, and generates reconstructed images from the output of the unrolled convolutional neural network by combining images from different shots. Each iteration of the unrolled network performs a first gradient update, applies the result to a first U-net in k-space, performs a second gradient update, and applies a second U-net in image space. The first gradient update and the second gradient update are based on a theoretical gradient from a physical measurement model.

Growing community of inventors

Stanford, CA, United States of America

Yuxin Hu