Neural Network Feature Map Fusion for Low-Overhead Image Restoration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The increasing calculation overheads in image restoration processes of neural networks pose a significant challenge, leading to inefficiencies in computer vision applications.
Innovation Solution
A neural network training method involving feature extraction using an adder neural network (AdderNet) with addition filtering processing, enhanced by a learnable power activation function, to maintain feature information and reduce energy consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional neural network models are used for image restoration, then image processing quality is improved, but calculation overheads and energy consumption increase
Solution Approach 1:
The patent segments the neural network into two distinct parts: an AdderNet module that performs efficient addition-based feature extraction, and a conventional neural network module that handles complex processing. This segmentation allows the system to benefit from both the energy efficiency of simple addition operations and the high processing quality of conventional networks, resolving the contradiction between energy consumption and image processing quality
Solution Approach 2:
The patent changes the operational parameters of the neural network by introducing a learnable scaling factor that dynamically adjusts the contribution of the AdderNet output. This parameter change enables the system to optimize the balance between computational efficiency and processing quality adaptively, reducing energy consumption while maintaining image restoration quality
2Use of energy by moving object
If addition filtering processing is used for feature extraction, then energy consumption is reduced, but feature information loss occurs
Solution Approach 1:
The patent introduces an intermediary mechanism where the AdderNet output serves as a complementary feature map that is added to the original input feature map. This intermediary approach ensures that no feature information is lost, as the original information is preserved and enhanced rather than replaced, while still benefiting from the energy-efficient addition operation
Solution Approach 2:
The patent uses copying by retaining the original input feature map and creating a copy that undergoes addition filtering processing. The processed copy is then combined with the original, ensuring that feature information is preserved while energy consumption is reduced through the efficient addition operation
3Measurement precision
If neural network complexity is increased to maintain feature information, then image processing quality is improved, but calculation overheads increase
Solution Approach 1:
The patent segments the computational workload into a simple AdderNet component and a conventional neural network component. This segmentation reduces overall calculation overheads by handling basic feature extraction through efficient addition operations, reserving complex computations only for where they are truly needed, thus maintaining image processing quality while reducing device complexity
Solution Approach 2:
The patent applies local quality by using addition filtering processing specifically for the feature extraction stage where simple operations suffice, while allowing conventional neural networks to handle more complex processing stages. This localized application of different processing strategies reduces overall calculation overheads while maintaining necessary image processing quality
Data Source
AI summary
A neural network training method, includes: obtaining an input feature map of a training image; performing feature extraction processing on the input feature map by using a feature extraction core of a neural network to obtain a first candidate feature map; adding the first candidate feature map and a second candidate feature map to obtain an output feature map, where the second candidate feature map is a feature map obtained after a value corresponding to each element in the input feature map is increased by N times, and N is greater than 0; determining an image processing result of the training image based on the output feature map; and adjusting a parameter of the neural network based on the image processing result.