UCLA researchers in the Department of Chemistry & Biochemistry have developed a novel image filtering algorithm that removes image noise while preserving image features with unprecedented fidelity.
Traditional image processing techniques use filtering algorithms based either on spatial smoothing (averaging) of pixel intensities or on distinguishing the true image signal from noise in a mathematically transformed domain (e.g. sorting image data by frequency content rather than by location). However, both of these approaches rely on imperfect assumptions about the statistical distribution of noise. Consequently, these methods often blur the image by suppressing certain spatial variations in intensity or by mistakenly discarding certain shapes as noise. The end result of these filtering techniques is noise reduction at the expense of diminished image quality.
UCLA researchers have developed an advanced image filtering algorithm that effectively removes image noise while preserving image features with unprecedented fidelity. This efficient de-noising algorithm employs a nonlinear filter based on multilayer perceptrons (MLPs) to groups of similar-looking image patches across multiple copies of the original image. This filtering technique outperforms current state-of-the-art noise removal algorithms including those based on collaborative filtering and total variation.
The described signal filtering algorithm has been implemented and tested on a set of MRI scans taken at UCLA. For all levels of initial image noise, images filtered using this novel algorithm were found superior to those filtered using other state-of-the-art de-noising methods, as quantified using the peak signal-to-noise ratio and the feature similarity index. Denoising of 23Na MRI images acquired at NYU (New York University) was also performed and demonstrated unprecedented performance for noise removal and feature preservation in extreme noise conditions.
The 23Na MRI images are unpublished yet, but available by request to the PI: email@example.com
|United States Of America||Issued Patent||9,953,246||04/24/2018||2015-240|
feature similarity index (FSIM), Image filtering, image restoration, image de-noising, signal processing, signal-to-noise ratio, noise removal, domain transformation, collaborative filtering (BM3D, BM4D), total variation filtering, nonlinear filtering, multi-layer perceptrons (MLP)