Multi-Branch Feature Map Processing for Low-Compute Visual Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks for machine tasks, particularly those involving visual tasks, require high computing power when used in combination with video or image encoding/decoding due to the need for reconstructing images from feature maps, which is computationally intensive.
Innovation Solution
A feature map processing method using a neural network model with multiple input adaptation branches and a post-processing part, where each branch downsamples feature maps at different ratios, ensuring the output size meets the post-processing part's requirements, thereby reducing computing power and maintaining analysis precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If an image domain task network is used to process visual tasks in combination with video or image encoding/decoding, then the network can process images, but high computing power is required due to the need to reconstruct images from feature maps
Solution Approach 1:
The patent extracts and removes the image reconstruction step from the processing pipeline. Instead of reconstructing images from feature maps and then processing them, the system directly processes the feature maps obtained from decoding, eliminating the computationally intensive reconstruction operation while maintaining the visual task processing capability
Solution Approach 2:
The patent inverts the conventional workflow by processing feature maps directly rather than processing reconstructed images. This inversion of the processing sequence (direct feature map processing instead of image reconstruction followed by processing) significantly reduces computing power requirements while achieving the same visual analysis objectives
2Device complexity
If feature maps with different resolutions are processed using a single neural network model, then the model structure is simple, but the analysis precision may be compromised for high-resolution feature maps
Solution Approach 1:
The patent segments the input adaptation part into multiple independent branches, each specialized for processing feature maps of specific resolution ranges. This segmentation allows each branch to optimize for its target resolution while the overall model structure remains unified and manageable
Solution Approach 2:
The patent creates a universal neural network model that can handle feature maps of multiple resolutions through the multi-branch input adaptation structure. Each branch serves a specific resolution range, but collectively they enable the single model to process diverse input resolutions without requiring separate specialized models for each resolution
3Measurement precision
If multiple independent neural network models are trained for different feature map resolutions, then each model can maintain high analysis precision for its target resolution, but the model deployment becomes complex and computationally expensive
Solution Approach 1:
The patent merges multiple resolution-specific processing paths into a single unified neural network model. Instead of deploying separate independent models for different resolutions, the system combines them into one model with shared post-processing parts, reducing deployment complexity while maintaining the precision benefits of resolution-specific processing
Solution Approach 2:
The patent designs a universal model architecture where a single model can handle multiple feature map resolutions through the multi-branch input adaptation structure. This eliminates the need for multiple separate models and their associated deployment complexities, while still providing optimized processing for each resolution range through the specialized branches
Data Source
AI summary
This application relates to the field of artificial intelligence technologies, and provides a feature map processing method and a related device. The method is implemented by invoking a neural network model that includes a plurality of input adaptation branches and a post-processing part. An output of each of the plurality of input adaptation branches is an input of the post-processing part, and each of the plurality of input adaptation branches can downsample a feature map at a different ratio.