# US 2024/0177279 A1 · Method and System for Deblurring Non-Uniformly Blurred Images

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2024/0177279 A1
- Document ID: US20240177279A1
- Status: Published application
- Kind: A1
- Application number: 18/520,796
- Filed: 2023-11-28
- Published: 2024-05-30
- Earliest claimed: 2022-11-28
- Assignees: Morgan State University
- Applicants: Morgan State University
- Inventors: Arlene Cole-Rhodes; Greig Richmond
- Web page: https://patentgrep.com/patent/US20240177279A1

## CPC

- G06T 5/73: Deblurring; Sharpening
- G06T 5/20: using local operators
- G06T 5/60: using machine learning, e.g. neural networks
- G06T 2207/20056: Discrete and fast Fourier transform, [DFT, FFT]
- G06T 2207/20081: Training; Learning
- G06T 2207/20084: Artificial neural networks [ANN]
- G06T: IMAGE DATA PROCESSING OR GENERATION, IN GENERAL

## Abstract

A patch-based deblurring method to deblur images blurred in a non-uniform manner (e.g., motion blur or camera shake) uses a modified total-variation (TV) minimization algorithm to make a network specifically designed to handle non-uniform blur.

## Claims

### Claim 1

A computer-implemented method for removing uniform or non-uniform blur from a blurred image, comprising:
a. using a processor to pass said blurred image through a first set of C 3D-convolutional filters trained on a neural network to identify image features that are characteristic of a high-resolution image and produce a first plurality of C filtered images in which potentially high-resolution features are emphasized;
b. using said processor to pass said first plurality of C filtered images through a second set of C 3D-convolutional filters trained on said neural network to identify said features to produce a second set of filtered images;
c. repeating step b with subsequent sets of 3D-convolutional filters trained on said neural network to identify said features to produce subsequent sets of filtered images until a final set of filtered images is produced following a predetermined (L) repetitions of step b;
d. using said processor to pass said final set of C filtered images (denoted Y1) to a first layer of said neural network, inputting to said first layer of said neural network a plurality of delta function filters (initial estimate of a blurring filters) (K0), and a filtered image initial estimate (G0), said first layer of said neural network generating a first set of revised estimates of the blurring filters (K1) and a first layer estimate (Z1) of said final set of filtered images (Y1), using an overlapping patch based approach,
i. wherein said first set of C revised estimates of the blurring filters (K1) correspond to different, but overlapping regions of said original image (Y), which are fused/assembled into a first layer blurring filter solution, and
ii. wherein said first layer of said neural network also fuses/assembles said first revised estimates of the first layer estimate (Z1) of said original image (Y);
e. using said processor to pass a penultimate set of C filtered images (Y2) to a second layer of said neural network, inputting to said second layer of said neural network said first set of revised estimates of the blurring filters (K1) and said first set of estimates (Z1), said second layer of said neural network generating a set of second revised estimates of the blurring filters (K2) and a second set of estimates (Z2) of said penultimate set of filtered images (Y2), using said overlapping patch based approach,
i. wherein said second set of revised estimates of the blurring filters (K2) correspond to said different but overlapping regions of said original image, and
ii. wherein said second layer of said neural network fuses/assembles said second revised estimates of the blurring filter into a second layer blurring filter solution (K2) and a second layer estimate (Z2) of said original images (Y);
f. repeating step e with subsequent layers of said neural network a number of times corresponding to said repetitions (L) of step b to produce a final blurring filter solutions (KL) and a final estimate (ZL) of said original images (Y);
g. using said processor to pass said blurred original image (Y), said first set of C 3D-convolutional filters, said final blurring filter solution (KL) and said final estimate (ZL) of said original blurred image (Y) through an inverse filter to produce a final deblurred image X.

### Claim 2 (depends on 1)

The computer-implemented method according to claim 1, wherein no additional information concerning the original blurred image is provided to said processor.

### Claim 3 (depends on 1)

The computer implemented method according to claim 1, wherein
said blurred image is processed by consecutive sets of C said 3D-convolution filters, said second set of C 3D-convolutional filters and said subsequent sets of C convolutional filters, where processed output from each said first set of 3D-convolution filters, said second set of 3D-convolutional filters and said subsequent sets of convolutional filters, is used as input to the next layer of said neural network,
a. wherein a combined number of said first set of 3D convolution filters, said second set of 3D-convolutional filters and said subsequent sets of convolutional filters, is equal to a total number of filters in the neural network and
b. wherein a last set of outputs from the L said subsequent sets of C convolutional filters is passed into said first layer of the neural network, an output from a penultimate set of convolutional filters is passed to the second layer of the neural network, and so forth until the output of the first set of convolutional filters is passed to the last layer of the neural network.
c. wherein C (i.e. eight to twenty-four) 3D-convolution filters make up the set in each layer and there are L (i.e. five to twenty) sets of filters, which correspond to five to twenty layers of the neural network respectively.
