Image Stylization via Feature Correction Maps

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image stylization methods often result in non-ideal output pictures when attempting to combine images with great visual differences, as the content of one image significantly affects the style of the output, and vice versa.

Innovation Solution

An image processing method using a convolutional neural network to extract feature maps from both the original and style reference images, calculate correction values based on Euclidean distances between feature blocks, and apply these corrections through a deconvolution process to maintain the content of the original image while adding the style of the reference image.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Manufacturing precision

If traditional image stylization methods are used to combine images with great visual differences, then the style transfer process is simple, but the content of one image significantly affects the style of the output and vice versa, resulting in non-ideal output pictures

Engineering Contradiction:
Improvestylization qualityVSAvoidprocessing complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the image processing into distinct feature extraction and feature combination stages using separate neural network components (convolutional network for content features, generative adversarial network for style features). This segmentation allows independent optimization of each stage, improving stylization quality while managing complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary feature map that mediates between content and style features. The feature map serves as a bridge that combines content features from the original image with style features from the reference image through controlled feature interaction, preventing direct contamination between content and style while achieving ideal stylization output.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Adaptability or versatility

If the content of the style reference picture is allowed to affect the output picture, then the style transfer is more natural, but the final output picture does not achieve ideal stylization效果

Engineering Contradiction:
Improvestyle adaptationVSAvoidstylization accuracy
Core Design Contradiction:
Adaptability or versatilityVSManufacturing precision

Solution Approach 1:

The patent applies local quality control by allowing style reference picture content to affect only specific regions of the output picture where style enhancement is needed, while preserving the original content structure through spatially selective feature combination. This is achieved through the feature map mechanism that locally integrates style features without globally contaminating content.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent dynamically adjusts the influence parameters of the style reference picture content based on regional needs. By controlling the weighting and integration strength of style features in different spatial locations and feature channels, the system achieves adaptability where style transfer is natural in appropriate regions while maintaining stylization accuracy in content-critical regions.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3770859B1Image processing method, image processing apparatus, and storage medium
Publication Date: 2021.11.03 BEIJING XIAOMI MOBILE SOFTWARE CO LTD
  • EP3770859B1 patent drawingFigure 1
  • EP3770859B1 patent drawingFigure 2
  • EP3770859B1 patent drawingFigure 3

AI summary

At least one embodiment of the present disclosure provides an image processing method, including: inputting (S11) an original image and a style reference image into a convolutional neural network to generate a first feature map of the original image and a second feature map of the style reference image; calculating (S12) a correction value of any pixel point in the first feature map based on the second feature map; replacing (S13) an original value of each pixel point in the first feature map with a corresponding corrected value to obtain a third feature map; and performing (S14) a deconvolution process on the third feature map to obtain a processed image. An apparatus and a storage medium are further provided.