Shared Feature Extraction Layers for Multi-Source Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing classification technologies, such as face recognition and vehicle plate recognition, face challenges in converging and performing well when combining multiple datasets due to cross-data issues and varying labelings, leading to poor performance, especially in scenarios with great gesture changes, low resolution, and poor quality data.
Innovation Solution
An apparatus and method that includes feature extraction and fusion units, along with a loss determination unit, which sets and shares feature extraction and fusion layers across training sets to align and fuse features correctly, allowing for accurate classification even across different datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple datasets are directly fused together to improve classification performance, then the classification performance can be effectively improved by utilizing the advantages of each dataset, but cross data exists among different datasets and the same data may have different labels, causing the trained classification model to not converge or have poor performances
Solution Approach 1:
The patent segments the feature processing pipeline by introducing separate feature extraction branches for each dataset, with each branch having its own feature extraction layers. This allows independent feature extraction while maintaining separate label spaces, preventing cross-data interference and enabling model convergence while still utilizing multiple datasets for improved classification performance.
Solution Approach 2:
The patent introduces an intermediary feature fusion mechanism that combines features from different dataset branches after independent extraction. This intermediary fusion layer integrates features from multiple sources while maintaining the integrity of each dataset's feature representation, allowing the model to leverage advantages of each dataset without direct conflict from cross-data labeling issues.
2Reliability
If more data is added to improve classification performance, then the classification technology achieves great progress, but collection of the large-scale data consumes a lot of time and energy
Solution Approach 1:
The patent creates a multi-functional classification model that can simultaneously process multiple datasets through shared feature extraction layers and unified training framework. This universal architecture allows the system to leverage existing published datasets (VGG2Face, Ms-Celeb-1M, MegaFace) without requiring new data collection, achieving improved classification performance by effectively utilizing advantages of each existing dataset through the proposed fusion mechanism.
3Reliability
If multiple datasets are used to improve classification performance, then the classification technology achieves great progress, but the device complexity increases due to multiple feature extraction and fusion layers
Solution Approach 1:
The patent merges multiple feature extraction branches into a unified model architecture with shared and private layers. The feature extraction layers for different datasets are merged in a way that allows sharing of common feature representations while maintaining dataset-specific capabilities. This merging approach reduces overall complexity compared to completely separate models while preserving the ability to handle cross-data issues through the fusion mechanism.
Data Source
AI summary
An apparatus for training a classification model includes: a feature extraction unit configured to set, with respect to each training set of a first predetermined number of training sets, feature extraction layers, and extract features of a sample image, where at least two of the training sets at least partially overlap; a feature fusion unit configured to set, with respect to training set, feature fusion layers, and perform a fusion on the extracted features of the sample image; and a loss determination unit configured to set, with respect to training set, a loss determination layer, calculate a loss function of the sample image based on the fused feature of the sample image, and train a classification model based on the loss function. The first predetermined number of training sets share at least one layer of feature fusion layers and feature extraction layers set with respect to each training set.