Adder Neural Network Training for Low-Energy Image Restoration

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The increasing calculation overheads in image restoration tasks within neural networks pose a significant challenge, particularly in computer vision applications, due to the high computational complexity of multiplication operations in convolutional neural networks.

Innovation Solution

Implementing an adder neural network (AdderNet) with addition filtering processing, utilizing L1 regular distances and learnable power activation functions to enhance high-frequency texture information, thereby reducing calculation amounts while preserving feature information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If convolutional neural networks use multiplication operations for feature extraction, then measurement precision is improved, but use of energy increases

Engineering Contradiction:
Improvefeature extraction precisionVSAvoidcomputation energy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent changes the mathematical operation parameter from multiplication to addition in the convolutional neural network. Specifically, it replaces the multiplication operation between input features and filter weights with addition operations, using L1-norm distance calculations instead. This parameter change maintains the ability to extract meaningful features while significantly reducing computational energy requirements.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent substitutes the traditional multiplication-based convolution operation with an addition-based operation. By replacing the multiplication mechanism with an addition mechanism (calculating L1-norm distances between input features and filter weights), the system achieves similar feature extraction functionality with lower computational complexity and energy consumption.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Use of energy by moving object

If adder neural network is used to reduce computation, then use of energy is reduced, but loss of information occurs

Engineering Contradiction:
Improvecomputation energy consumptionVSAvoidfeature information loss
Core Design Contradiction:
Use of energy by moving objectVSLoss of information

Solution Approach 1:

The patent introduces feedback mechanisms through learnable power activation functions that adaptively adjust the processing of feature maps. The network uses backpropagation to learn optimal parameters for the power activation function, ensuring that the addition-based operation preserves essential feature information while reducing computation. The feedback loop allows the network to compensate for potential information loss through learned parameter adjustments.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent makes the addition operation dynamic by introducing learnable power activation functions with adjustable parameters. Instead of a static addition operation, the network dynamically adjusts the transformation applied to feature maps based on learned parameters, allowing it to preserve important feature information while maintaining the computational efficiency of addition-based operations.

Inventive Principle:
Principle #15Dynamics

3Manufacturing precision

If traditional convolution operation is used, then manufacturing precision is maintained, but productivity decreases

Engineering Contradiction:
Improveimage restoration qualityVSAvoidprocessing speed
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The patent changes the operational parameter from multiplication to addition in the convolution operation. This parameter change reduces the computational complexity from O(n) multiplications per element to O(n) additions per element, significantly improving processing speed while maintaining image restoration quality through the use of L1-norm distance calculations and learnable activation functions.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent substitutes the traditional multiplication-based convolution mechanism with an addition-based mechanism. By replacing multiplication with addition operations (calculating L1-norm distances), the system achieves faster processing speeds due to the simpler computational nature of addition, while maintaining manufacturing precision through careful design of the addition-based feature extraction process.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4163832B1Neural network training method and apparatus, and image processing method and apparatus
Publication Date: 2026.03.18 HUAWEI TECH CO LTD
  • EP4163832B1 patent drawingFigure 1
  • EP4163832B1 patent drawingFigure 2
  • EP4163832B1 patent drawingFigure 3

AI summary

A neural network training method, an image processing method, and an apparatus are provided. The method relates to the field of artificial intelligence, and specifically, to the field of computer vision. The method includes: obtaining an input feature map of a training image (S510); 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 (S520); adding the first candidate feature map and a second candidate feature map to obtain an output feature map (S530), 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 (S540); and adjusting a parameter of the neural network based on the image processing result. According to the method, energy consumption of a neural network model can be effectively reduced without affecting image processing effect (S550).