Sequential Convolutional Neural Network for Image Deblurring

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing methods using parallel convolutional neural networks for deblurring images suffer from weak information exchange between networks, leading to increased computational requirements for producing clear images.

Innovation Solution

The proposed method involves training a single convolutional neural network in a sequential manner, with each stage enhancing the network's capability to deblur images, include motion differences, and recover hidden information, thereby reducing computational load.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If multiple convolutional neural networks are arranged in parallel to process blurred images, then multiple clear images can be obtained, but the computational complexity increases and information exchange between networks is weak

Engineering Contradiction:
Improvenumber of clear images outputVSAvoidcomputational complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent merges multiple parallel convolutional neural networks into a single multi-branch network architecture. The blurred image is simultaneously input into multiple branches (first branch for deblurring, second branch for motion estimation, third branch for detail recovery), which then interact and exchange information through shared layers and feedback mechanisms, finally outputting multiple clear images with different motion characteristics. This combining approach maintains the ability to produce multiple clear images while reducing overall computational complexity and enabling information exchange between previously isolated networks.

Inventive Principle:
Principle #5Merging (Combining)

2Productivity

If a larger quantity of convolutional neural networks are trained to produce more clear pictures, then the output quantity increases, but the computational burden increases

Engineering Contradiction:
Improvequantity of clear picturesVSAvoidcomputational burden
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

The patent designs a multi-functional single network that can simultaneously perform multiple functions: deblurring (first branch), motion estimation (second branch), and detail recovery (third branch). This universal network processes the blurred image through different branches to produce multiple clear images with various motion characteristics, replacing the need for multiple separate specialized networks. The shared computational resources and parameter reuse across branches significantly reduce the computational burden while maintaining high productivity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Productivity

If parallel convolutional neural networks are used for image deblurring, then processing capability is distributed, but information exchange capability between networks is weak

Engineering Contradiction:
Improveprocessing capabilityVSAvoidinformation exchange capability
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent implements feedback mechanisms within the multi-branch network architecture. The motion estimation branch (second branch) provides motion information back to the deblurring branch (first branch), and the detail recovery branch (third branch) refines the output by incorporating information from previous branches. This feedback loop enables continuous information exchange and refinement across what would otherwise be isolated processing paths, ensuring that each branch benefits from the computational results of others while maintaining distributed processing capability.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3937123B1Image processing method and apparatus, and computer device and storage medium
Publication Date: 2025.05.07 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3937123B1 patent drawingFigure 1~2
  • EP3937123B1 patent drawingFigure 3
  • EP3937123B1 patent drawingFigure 4

AI summary

An image processing method. The method comprises: training an initial network to obtain a first network; training the first network to obtain a second network; training the second network to obtain a third network; and upon receiving an image processing instruction, performing image processing on an original image according to the third network to obtain a plurality of target images.