Feature Fusion in Adder Neural Networks for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks with adder operations struggle to achieve effective object detection due to high power consumption and sparsity of high-level features, making them unsuitable for tasks like object detection.
Innovation Solution
The method involves fusing low-level features to high-level features using an adder neural network with a target network that includes multiple branches for feature extraction, enhancing the network's expression capability through skip connections and gradient calculation, improving object detection precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If convolutional neural network with multiplication operations is used, then object detection precision is improved, but power consumption increases
Solution Approach 1:
The patent replaces multiplication operations with addition operations in the neural network computation. Specifically, it uses an adder neural network where the activation function is replaced by an addition operation, substituting the traditional multiplication-based convolutional neural network with an addition-based approach that consumes less power while maintaining detection capability
Solution Approach 2:
The patent changes the computational parameter from multiplication to addition. By modifying the activation function from a traditional nonlinear activation (like ReLU or sigmoid) to a simple addition operation, the network achieves lower power consumption. The patent also adjusts the network structure by adding multiple branches to compensate for the reduced computational capability
2Use of energy by moving object
If adder neural network is used to reduce power consumption, then power consumption is reduced, but object detection precision deteriorates
Solution Approach 1:
The patent divides the neural network into multiple branches, where each branch processes features at different levels. The network is segmented into a first branch for low-level feature processing and a second branch for high-level feature processing, allowing the adder neural network to maintain detection precision by distributing computational tasks across specialized pathways
Solution Approach 2:
The patent adds a new dimension to the network architecture by introducing multiple parallel branches that process features at different hierarchical levels. This dimensional expansion allows the network to compensate for the simplified addition operation by processing information through multiple pathways, thereby maintaining detection precision despite the reduced computational complexity
3Loss of information
If high-level feature is extracted, then semantic information is improved, but feature sparseness increases
Solution Approach 1:
The patent merges low-level features and high-level features through a feature fusion mechanism. By combining the dense low-level features (which have rich spatial information) with the semantic high-level features (which have rich conceptual information), the network achieves a balanced feature representation that maintains both semantic meaning and feature density
Solution Approach 2:
The patent implements a nested feature structure where low-level features are nested within the processing pipeline and fused with high-level features. The network structure allows low-level features to be processed and integrated into the high-level feature representation, creating a nested hierarchy where each level contributes to the final detection result
Data Source
AI summary
An image processing method is disclosed in embodiments of this disclosure and is applied to the field of artificial intelligence. The method includes: obtaining an input feature map of an image to be processed, where the input feature map includes a first input sub-feature map and a second input sub-feature map, and resolution of the first input sub-feature map is higher than resolution of the second input sub-feature map; performing feature fusion processing on the input feature map by using a target network, to obtain an output feature map, where a feature of the first input sub-feature map is fused to a feature of the second input sub-feature map from a low level to a high level in the target network; and performing, based on the output feature map, object detection on the image to be processed, to obtain an object detection result.