Nonlinear Image Compression Processing for Receptive Field Expansion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image/video compression algorithms face challenges in optimizing the non-linear transformation units, which affect the rate-distortion performance and computing efficiency, limiting the effectiveness of deep learning-based compression methods.
Innovation Solution
A non-linear processing method involving multiple stages of convolution processing and element-wise multiplication operations is employed to expand the receptive field and reduce calculation amounts, enhancing local attention and balancing computing power while maintaining model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If a single non-linear transformation unit is used in image compression, then the model structure is simple, but the receptive field is limited and rate-distortion performance is suboptimal
Solution Approach 1:
The patent divides a single complex non-linear transformation unit into multiple simpler non-linear transformation units connected in sequence. Each unit performs partial processing with its own convolution operations and activation functions, collectively achieving a larger effective receptive field and improved rate-distortion performance while keeping individual unit structures simple and manageable
2Manufacturing precision
If multiple convolution operations are performed sequentially, then the receptive field is expanded and local attention is enhanced, but the calculation amount increases
Solution Approach 1:
The patent segments the convolutional processing into multiple stages, where each non-linear transformation unit performs a subset of convolution operations. This segmentation allows the receptive field to be gradually expanded through sequential processing while distributing the computational load across multiple simpler units, improving both local attention capability and calculation efficiency
Solution Approach 2:
The patent employs dynamic feature processing where each non-linear transformation unit adaptively processes features from the previous unit. The dynamic nature of feature transformation at each stage allows the system to focus computational resources on important local features, enhancing local attention while optimizing overall calculation efficiency
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This application provides a non-linear processing method and apparatus for image compression. The non-linear processing method for image compression in this application includes: obtaining a to-be-processed first image feature; and performing non-linear transformation processing on the first image feature to obtain a processed image feature, where the non-linear transformation processing sequentially includes first convolution processing, second convolution processing, and an element-wise multiplication operation. This application can expand a receptive field, reduce a calculation amount, further enhance local attention, maintain model performance, and effectively balance computing power.