Image Super-Resolution via Gradient Feature Extraction and Highway Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image super-resolution methods, particularly those based on convolutional neural networks (CNNs), face challenges in achieving high-quality reconstructions due to convergence issues and excessive computational requirements, leading to suboptimal super-resolution effects.
Innovation Solution
The proposed method involves preprocessing low-resolution images to obtain vertical and horizontal gradient maps and a luminance map, which are then used to create a feature map for input into a deep neural network. This network performs nonlinear transformations and weighted processing, utilizing a highway network structure to overcome convergence problems and improve image quality by directly connecting certain layers, thereby reducing the number of parameters and enhancing processing speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If the network depth is increased to achieve better super-resolution effect, then the super-resolution quality is improved, but the quantity of parameters becomes excessively large and the processing time increases
Solution Approach 1:
The patent extracts and utilizes gradient information (vertical gradient map, horizontal gradient map) and luminance map from the low-resolution image as separate feature dimensions. This extraction approach allows the network to focus on key structural and intensity information, achieving better super-resolution quality without requiring excessive network depth and parameters
Solution Approach 2:
The patent transforms the input low-resolution image into a multi-dimensional feature space by creating vertical gradient map, horizontal gradient map, and luminance map as three different dimensions. This dimensional transformation enriches the feature representation, enabling the network to achieve better reconstruction quality with reduced computational complexity
2Manufacturing precision
If the network depth is increased to achieve better super-resolution effect, then the super-resolution quality is improved, but the processing speed becomes slower and time consumption increases
Solution Approach 1:
The patent extracts gradient information and luminance information as separate feature maps before inputting to the network. This preprocessing extraction reduces the computational burden during network processing, improving processing speed while maintaining the ability to achieve high-quality super-resolution
Solution Approach 2:
The patent performs preliminary processing to generate vertical gradient map, horizontal gradient map, and luminance map before the main super-resolution processing. This preliminary action prepares the data in an optimized format, reducing the computational complexity and time required during the actual network processing stage
3Productivity
If conventional interpolation methods are used for super-resolution, then the processing is simple and fast, but the reconstructed high-resolution image quality is poor
Solution Approach 1:
The patent introduces gradient maps and luminance maps as intermediary feature representations between the low-resolution input and the high-resolution output. These intermediaries capture essential structural and intensity information, enabling quality reconstruction while maintaining processing efficiency through the use of gradient-based features
Data Source
AI summary
An image super-resolution method includes preprocessing the low-resolution image to obtain a vertical gradient map, a horizontal gradient map, and a luminance map, which are used as three different dimensions of information to constitute a to-be-input feature map, performing size conversion on the to-be-input feature map to obtain an input feature map, performing nonlinear transformation on the input feature map to obtain an input feature map obtained after the nonlinear transformation, and performing weighted processing on the input feature map and the input feature map obtained after the nonlinear transformation, to obtain an output feature map, performing size conversion on the output feature map to obtain a residual map, and combining the low-resolution image and the residual map to obtain a high-resolution image.