Multi-Size Sparse Convolution Kernels for Efficient Image Super-Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing super-resolution methods based on neural networks face challenges with receptive field expansion, leading to increased parameters and high computational requirements, which limits their application.
Innovation Solution
The method determines convolution kernels of multiple sizes for feature extraction, applies sparsity constraints using a preset objective function, and performs feature extraction on images using these constrained kernels.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If a large-size convolution kernel is adopted for receptive field expansion, then the receptive field is expanded, but the parameters increase in a square number and high requirements for network computing power are introduced
Solution Approach 1:
The patent divides a large convolution kernel into multiple small convolution kernels arranged in a group. Instead of using one large kernel that would require O(n²) parameters, the patent segments it into multiple small kernels (e.g., three 3×3 kernels) that collectively cover the same receptive field area. This segmentation reduces the total parameter count from quadratic to linear growth with respect to receptive field size, resolving the contradiction between receptive field expansion and parameter complexity.
2Area of stationary object
If multiple-layer convolution kernels are used for receptive field expansion with small-size kernels, then the receptive field is expanded, but the difficulty of neural network training increases
Solution Approach 1:
The patent merges multiple small convolution kernels into a single grouped convolution layer that operates in parallel. Instead of stacking multiple sequential convolution layers (which increases training difficulty due to deeper optimization paths), the patent combines multiple kernels of the same size into one grouped layer. This merging approach maintains the expanded receptive field while keeping the network depth unchanged, thereby reducing training difficulty while achieving the desired receptive field size.
3Manufacturing precision
If a deeper network structure is adopted, then the receptive field is expanded and image reconstruction effect is improved, but the burden of network computing is introduced
Solution Approach 1:
The patent transitions from expanding the receptive field through the depth dimension (adding more layers) to expanding it through the width dimension (using grouped convolutions with multiple kernels in parallel). This dimensional shift allows the network to achieve the same receptive field coverage without increasing depth, thereby maintaining image reconstruction quality while significantly reducing computational burden and energy consumption.
Data Source
AI summary
The present application provides an image processing method, an image processing device, a computer storage medium and a terminal, the image processing method includes: determining convolution kernels of at least two sizes for feature extraction; performing sparsity constraint for the determined convolution kernels of at least two sizes for feature extraction through a preset objective function; and performing feature extraction on an image based on the convolution kernels subjected to the sparsity constraint.