Method and apparatus with object classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object classification methods in neural networks face challenges in achieving high accuracy while minimizing resource consumption, particularly in mobile devices with limited memory and power, due to the loss of local information and increased computational demands.
Innovation Solution
A method and apparatus that merge global and local feature data using a neural network, employing a hierarchical structure and cross-attention feature aggregation to enhance accuracy while reducing resource usage, by integrating feature data from multiple layers and applying weights to optimize dimensionality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If global feature data from upper layers is used for classification, then classification accuracy is improved, but local information is lost
Solution Approach 1:
The patent merges global feature data from upper layers with local feature data from lower layers to create comprehensive feature representations. This combining approach preserves both global context and local details simultaneously, resolving the contradiction between using global features for accuracy and maintaining local information.
Solution Approach 2:
The patent introduces a multi-dimensional feature aggregation approach where features are processed across different spatial and hierarchical dimensions. By operating in multiple dimensions rather than a single global dimension, the system maintains local information while still benefiting from global context.
2Measurement precision
If multiple feature extraction layers are used to improve accuracy, then classification performance is enhanced, but computational demand and resource consumption increase
Solution Approach 1:
The patent extracts and aggregates only the most relevant feature data from multiple layers rather than processing all features from all layers. This selective extraction reduces computational demand while maintaining the accuracy benefits of multi-layer feature extraction.
Solution Approach 2:
The patent applies partial action by using a subset of feature data from upper layers (specifically, selectively aggregated features) rather than processing all features. This partial processing approach reduces computational overhead while still achieving improved classification accuracy through the inclusion of global context.
3Measurement precision
If feature data from multiple layers is aggregated, then classification accuracy is improved, but memory requirements increase
Solution Approach 1:
The patent extracts and stores only the essential aggregated feature data from multiple layers rather than maintaining all intermediate feature representations. This selective storage approach reduces memory requirements while preserving the accuracy benefits of multi-layer feature aggregation.
Data Source
AI summary
An object classification method and apparatus are disclosed. The object classification method includes receiving an input image, storing first feature data extracted by a first feature extraction layer of a neural network configured to extract features of the input image, receiving second feature data from a second feature extraction layer which is an upper layer of the first feature extraction layer, generating merged feature data by merging the first feature data and the second feature data, and classifying an object in the input image based on the merged feature data.