Prototype learning and clustering based small sample unknown environment sound discovery method

By using a prototype learning and clustering approach, log-Mel spectrum features are extracted and combined with a feature extractor and a clusterer, solving the problem of discovering unknown categories in environmental sound recognition under small sample conditions, and achieving high-precision and adaptive environmental sound monitoring.

CN122135737APending Publication Date: 2026-06-02SOUTH CHINA UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
SOUTH CHINA UNIV OF TECH
Filing Date
2026-02-03
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively identify unknown categories in environmental sounds under small sample conditions, and traditional methods lack sufficient accuracy in complex acoustic environments, failing to adapt to dynamic changes and the scarcity of labeled data.

Method used

A small-sample unknown environmental sound discovery method based on prototype learning and clustering is adopted. By extracting log-Mel spectrum features and combining a feature extractor, classifier and clusterer, the method utilizes the similarity of known class prototypes and the mapping of unknown class clusters to achieve open set recognition and unknown category discovery.

Benefits of technology

It significantly improves the recognition capability under small sample conditions, realizes accurate automatic discovery and dynamic expansion of unknown categories, and enhances the system's adaptability and robustness in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122135737A_ABST
    Figure CN122135737A_ABST
Patent Text Reader

Abstract

This invention discloses a method for discovering unknown environmental sounds in small samples based on prototype learning and clustering. The method first extracts log-Mel spectrum features from environmental sound samples and inputs them into a feature extractor to obtain the features of the environmental sound samples to be detected. These features are then input into a classifier for initial classification. Based on the prototype similarity of environmental sound samples of known classes, it is determined whether the features of the environmental sound samples to be detected belong to a known or unknown class. Environmental sound samples classified as belonging to a known class undergo a second classification into corresponding known subclasses. Environmental sound samples classified as belonging to an unknown class are input into a clusterer, and based on clusters of unknown class environmental sound samples, they are clustered into unknown subclasses. Finally, based on the features of the unknown class environmental sound samples and the responsibility value of each unknown class environmental sound sample cluster, they are classified into corresponding unknown subclasses. This method identifies old classes and discovers new classes of environmental sounds based on a small number of training samples, and is applicable to scenarios such as autonomous driving assistance and ecological audio monitoring.
Need to check novelty before this filing date? Find Prior Art