Feature Learning System for Human Action Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for learning human actions in images face challenges in distinguishing between 'totally different' and 'similar but different' actions, leading to reduced identification precision and increased time requirements for learning and identification processes.
Innovation Solution
A feature learning system that defines a degree of similarity between action feature vectors, generates learning data incorporating these similarities, and performs machine learning using this data to stabilize the learning process and improve identification efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional metric learning handles both 'totally different actions' and 'similar but different actions' as 'different actions', then the learning process is simple, but identification precision deteriorates due to conversion exaggerating irrelevant differences
Solution Approach 1:
The patent applies local quality by differentiating the handling of action pairs based on their similarity characteristics. Instead of uniform treatment, the system identifies and processes 'similar but different actions' separately from 'totally different actions', applying appropriate learning strategies to each group to prevent conversion exaggeration of irrelevant differences while maintaining learning efficiency.
Solution Approach 2:
The patent segments the action classification task into multiple sub-tasks by introducing intermediate classification layers. The system divides the learning process into stages where actions are first classified into broader categories, then into more specific sub-categories, allowing nuanced differentiation between similar but different actions without overwhelming complexity in a single learning step.
2Measurement precision
If actions are classified into multiple topics with separate discriminators, then identification considering similarity is enabled, but learning and identification time increases due to repeated processing
Solution Approach 1:
The patent merges multiple classification tasks into a unified learning framework. Instead of training separate discriminators for each topic and performing multiple identification passes, the system integrates similarity-based classification into a single learning process that simultaneously handles both broad category identification and fine-grained differentiation, reducing redundant computations.
Solution Approach 2:
The patent performs preliminary classification actions during the learning phase to pre-organize action data into hierarchical structures. By pre-establishing topic categories and their relationships during training, the system eliminates the need for repeated classification passes during identification, significantly reducing time consumption while maintaining precision.
Data Source
AI summary
A feature learning system (100) includes a similarity definition unit (101), a learning data generation unit (102), and a learning unit (103). The similarity definition unit (101) defines a degree of similarity between two classes related to two feature vectors, respectively. The learning data generation unit (102) acquires the degree of similarity, based on a combination of classes to which a plurality of feature vectors acquired as processing targets belong, respectively, and generates learning data including the plurality of feature vectors and the degree of similarity. The learning unit (103) performs machine learning using the learning data.