Semi-supervised audio event identification method based on depth mutual information maximization

Abstract

The invention relates to a semi-supervised audio event identification method based on depth mutual information maximization. A semi-supervised neural network model is used as a backbone, a depth mutual information maximization consistency-based regular constraint and a cross entropy classification constraint are designed, a semi-supervised learning model is constructed, a mutual information discriminator is designed to estimate mutual information between deep representation vectors of the model, the model mines potential relations between samples through global mutual information so as to enhance consistency and nonlinear correlation between global representations, and a semi-supervised audio event classification model with high robustness is obtained; and neural network model parameters are optimized by using a gradient descent method, and the audio event samples are classified. The method has the advantages of being small in error, high in robustness, high in precision and the like,the requirement for sound event classification can be met under the condition that label data is insufficient, and high application value is achieved.

Description

Technical field: [0001] The invention relates to an audio event recognition method, in particular to a semi-supervised audio event recognition method based on depth mutual information maximization. Background technique: [0002] Audio signals carry a wealth of information about the everyday environment and where physical events occur. Humans can easily perceive the sound scene they are in (busy street, office, etc.) and recognize individual audio events (cars, footsteps, etc.). Automatic detection of audio events has many real-life applications. For traditional sound event classification, it is more dependent on artificial preprocessing features, such as manually selecting the number of filters of MFCC, pitch centroid feature energy, etc. These traditional methods lack efficiency and practicality in current applications. Sound event classification methods based on deep learning use neural networks for automatic feature extraction and result classification, but current sta...

AI Technical Summary

Problems solved by technology

[0004] In order to overcome the randomness and one-sidedness of the constraints of the consistency regularization method in the traditional semi-supervised audio event recognition method, it cannot guide the model to mine the most effective internal representation

The invention provides a semi-supervised audio event recognition method based on the maximization of depth mutual information. The method uses the representation vector inside the model to impose consistency constraints on the model with the goal of maximizing the representation mutual information, and strengthens the recognition of the representation vector for the same category. The nonlinear statistical correlation between data solves the problem that traditional consistency regularization methods cannot guide the model to mine the most effective internal representation, and improves the robustness of modeling

Images