CNN Image Distortion Correction via Side Information Components

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional image distortion elimination methods require complex filtering processes and high bitrates due to manual design and adaptation of filter structures, limiting their efficiency and flexibility, especially in handling images with varying distortion degrees.

Innovation Solution

A CNN-based method that generates side information components representing distortion features of a distorted image, which are then input into a pre-trained convolutional neural network for convolution filtering to eliminate distortion, allowing for data-driven automatic learning and improved generalization across different distortion levels.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If conventional adaptive filters (SAO, ALF) are used to eliminate image distortion, then filtering quality is improved, but device complexity and bitrate increase due to manual design of filter structures and configuration of filter coefficients based on local statistical information

Engineering Contradiction:

Improveimage distortion elimination qualityVSAvoidfilter structure complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent replaces manually designed mechanical filtering systems with a data-driven deep learning system. A convolutional neural network is trained to automatically learn optimal filtering operations from training data, substituting the manual analysis and artificial filter design process with an automated neural network that adapts to different distortion characteristics without requiring explicit programming of filter structures or coefficients

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the approach from manually configuring filter parameters based on statistical analysis to having the neural network automatically learn and adapt parameters during training. The network dynamically adjusts its internal parameters (weights and biases) based on the input image characteristics, eliminating the need for manual parameter configuration and reducing device complexity

Inventive Principle:

Principle #35Parameter changes

2Manufacturing precision

If adaptive filters with local statistical information are used, then distortion elimination performance is improved, but the number of encoded bits increases due to writing filter parameters into bitstream

Engineering Contradiction:

Improvedistortion elimination performanceVSAvoidnumber of encoded bits

Core Design Contradiction:

Manufacturing precisionVSLoss of information

Solution Approach 1:

The patent extracts only the essential distortion characteristics needed for filtering by using a lightweight feature extraction module that identifies key distortion patterns without requiring full statistical analysis. This extraction approach reduces the amount of information that needs to be encoded and transmitted while maintaining sufficient data for effective distortion elimination

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent uses a pre-trained neural network model that has already learned optimal filtering strategies from extensive training data. Instead of encoding and transmitting detailed filter parameters and statistical information, the system copies the learned knowledge into a compact network structure that can be deployed with minimal bitstream overhead, achieving high performance without increasing encoded bit count

Inventive Principle:

Principle #26Copying

3Ease of manufacture

If manual feature analysis and artificial filter design are used, then filtering can be performed, but flexibility and adaptability to different distortion types are reduced

Engineering Contradiction:

Improvefiltering implementationVSAvoidadaptability to different distortion types

Core Design Contradiction:

Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The patent implements a dynamic filtering system where the neural network adapts its behavior based on the input image characteristics. The network dynamically selects and adjusts filtering operations according to the specific distortion patterns detected in each image, providing high flexibility and adaptability without requiring manual redesign for different distortion types. This dynamic approach contrasts with static manually-designed filters that must be reconfigured for different scenarios

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS11250546B2Image distortion correction method and device and electronic device

Publication Date: 2022.02.15 HANGZHOU HIKVISION DIGITAL TECHNOLOGY CO LTD

AI summary

Embodiments of the present application provide a method, apparatus, and electronic device for eliminating distortion of a distorted image. Side information components of a distorted image are generated. The distorted image is resulted from image processing on an original image. Side information components represent distortion features of the distorted image with respect to original image. Distorted image color components of the distorted image and the side information components are input into a pre-established convolutional neural network model for convolution filtering to obtain distortion-eliminated image color components. The convolutional neural network model is obtained through training based on a preset training set. The training set comprises original sample images, distorted image color components of multiple distorted images corresponding to each of the original sample images, and side information components of each of the distorted images. Thereby, a CNN-based distortion elimination process different from related art is provided.