Image Feature Rearrangement for Semantic Texture Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing methods, including those using artificial intelligence models, fail to effectively recover details in large areas of texture missing, generate unreasonable textures due to lack of explicit semantic information, and suffer from over-processing without adequate judgment of image quality degradation, particularly in low-light conditions.
Innovation Solution
An image processing method that rearranges feature blocks in a feature space, uses a machine learning model with a semantic encoding branch and quality degradation estimation branch to guide texture generation, and combines feature blocks based on importance and correlation, ensuring reasonable texture recovery and preventing over-processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used for object detection, then detection accuracy is improved, but computational resources and processing time are increased
Solution Approach 1:
The patent divides the image processing task into multiple stages: a lightweight detection model processes the entire image to identify potential objects, while a separate deep learning model only processes regions of interest identified by the lightweight model. This segmentation of processing tasks reduces overall computational resource consumption while maintaining high detection accuracy.
Solution Approach 2:
The patent applies partial action by using a lightweight detection model to perform initial screening of the entire image, then selectively applying the computationally intensive deep learning model only to specific regions where objects are detected. This partial application of heavy computation only where needed reduces total computational resource usage while maintaining accuracy.
2Measurement precision
If deep learning models are used for object detection, then detection accuracy is improved, but processing time is increased
Solution Approach 1:
The patent segments the detection process into two stages: a fast lightweight model performs initial detection across the entire image quickly, then a deeper model processes only the identified regions of interest. This segmentation reduces total processing time compared to applying the deep learning model to the entire image.
Solution Approach 2:
The lightweight detection model performs preliminary action by quickly scanning the entire image and identifying potential object regions before the more time-consuming deep learning model processes those specific regions. This preliminary screening action reduces the overall processing time by avoiding unnecessary deep computation in empty areas.
3Measurement precision
If multiple detection models are used, then detection accuracy is improved, but device complexity is increased
Solution Approach 1:
The patent segments the detection system into a lightweight model for initial processing and a deep learning model for detailed analysis of specific regions. This segmentation allows the system to achieve high accuracy through model specialization rather than using a single complex model, thereby managing system complexity.
Solution Approach 2:
The lightweight detection model acts as an intermediary that bridges the simple input image and the complex deep learning model. It processes the image first, identifies relevant regions, and passes only those regions to the deep learning model, thereby managing system complexity by controlling the flow of data between models.
Data Source
Figure 1~3
Figure 4
Figure 5
AI summary
The present disclosure relates to an image processing method and device, an electronic apparatus and a storage medium, and the image processing method includes: acquiring an input image; detecting a target area in the input image; and processing the target area, wherein the processing of the target area includes: obtaining a feature map of the target area, rearranging feature blocks in the feature map in a feature space, and obtaining an output image after the target area is processed based on the rearranged feature blocks and the feature map.