CNN Super Resolution with Separate Channels to Reduce Image Artifacts

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing deep learning based super-resolution techniques suffer from artifacts such as overshoot, undershoot, and aliasing, particularly in screen and gaming content, due to corruption of high frequency information during downsampling, and conventional methods introduce high computational workload and blurriness.

Innovation Solution

A scalable neural network approach using nearest neighbor downsampling, self-similarity loss, and PReLU activation to process illuminance and chrominance channels separately, reducing computational resources and eliminating artifacts while maintaining sharpness.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional deep learning based super-resolution techniques are used, then high resolution images can be generated, but artifacts such as overshoot, undershoot, and aliasing occur particularly in screen and gaming content

Engineering Contradiction:

Improveimage resolutionVSAvoidartifacts (overshoot, undershoot, aliasing)

Core Design Contradiction:

Measurement precisionVSObject-affected harmful factors

Solution Approach 1:

The patent segments the image processing into two distinct pathways: a first convolutional neural network processes illuminance components while a second convolutional neural network processes chrominance components. This segmentation allows each network to be optimized for its specific component type, preventing the artifacts that occur when treating all image data uniformly.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies different processing qualities and methods to different components of the image. Illuminance components (brightness information) are processed through one neural network with specific architecture, while chrominance components (color information) are processed through another neural network. This local quality approach ensures that each component receives the appropriate processing to avoid artifacts.

Inventive Principle:

Principle #3Local quality

2Measurement precision

If conventional super-resolution methods are used, then image resolution is improved, but computational workload increases significantly

Engineering Contradiction:

Improveimage resolutionVSAvoidcomputational workload

Core Design Contradiction:

Measurement precisionVSPower

Solution Approach 1:

By segmenting the processing into two separate neural networks (one for illuminance, one for chrominance), the computational workload is divided and can be processed in parallel. This reduces the overall computational burden compared to processing the entire image through a single large network.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies super-resolution processing selectively to different components rather than treating the entire image uniformly. The illuminance and chrominance components are processed separately with appropriate computational resources allocated to each, avoiding excessive computation on components that require less processing.

Inventive Principle:

Principle #16Partial or excessive action

3Measurement precision

If conventional super-resolution methods are used, then image resolution is enhanced, but blurriness is introduced

Engineering Contradiction:

Improveimage resolutionVSAvoidimage sharpness

Core Design Contradiction:

Measurement precisionVSManufacturing precision

Solution Approach 1:

The patent applies different processing approaches to different image components to maintain local quality. Illuminance components are processed through a first CNN optimized for brightness information, while chrominance components are processed through a second CNN optimized for color information. This ensures that sharpness is maintained in critical areas without introducing blurriness.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent uses nearest-neighbor downsampling to create training pairs that preserve sharp edges and high-frequency information. This copying approach maintains the original image's sharp characteristics during the training process, enabling the neural networks to learn artifact-free super-resolution.

Inventive Principle:

Principle #26Copying

4Ease of manufacture

If conventional downsampling techniques are used during training, then training data is prepared, but high frequency information is corrupted causing artifacts in screen and gaming content

Engineering Contradiction:

Improvetraining data preparationVSAvoidhigh frequency information

Core Design Contradiction:

Ease of manufactureVSLoss of information

Solution Approach 1:

The patent uses nearest-neighbor downsampling to create training pairs by copying and repeating pixel values rather than using interpolation methods. This copying approach preserves sharp edges and high-frequency information in the training data, preventing the corruption that occurs with conventional downsampling techniques and enabling artifact-free super-resolution.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS12367547B2Super resolution using convolutional neural network

Publication Date: 2025.07.22 INTEL CORP

AI summary

An apparatus for super resolution imaging includes a convolutional neural network (104) to receive a low resolution frame (102) and generate a high resolution illuminance component frame. The apparatus also includes a hardware scaler (106) to receive the low resolution frame (102) and generate a second high resolution chrominance component frame. The apparatus further includes a combiner (108) to combine the high resolution illuminance component frame and the high resolution chrominance component frame to generate a high resolution frame (110).