AI Sound Source Mapping for High-Resolution Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for obtaining sound source maps using deep learning face challenges in achieving high spatial resolution, short computation time, and high accuracy, particularly when the number of sound sources is unknown or exceeds a predetermined maximum.
Innovation Solution
A method and system that generate a target map for training a neural network, allowing for the generation and synthesis of sound source maps without grid constraints, using a target map inference and analysis process to enhance accuracy and reduce computation time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deconvolution methods are used to improve spatial resolution, then spatial resolution is improved, but computation time becomes too long
Solution Approach 1:
The patent pre-calculates and stores point spread functions (PSFs) at different grid positions before actual sound source localization. This preliminary action allows the system to quickly retrieve and apply appropriate PSFs during inference without performing repetitive deconvolution calculations, thus achieving high spatial resolution with reduced computation time
Solution Approach 2:
The patent pre-generates target maps with known sound source positions and strengths to train the neural network beforehand. This training phase enables the network to learn the mapping from beamforming maps to accurate sound source parameters, allowing fast and accurate inference without complex real-time calculations
2Productivity
If grid-based deep learning methods are used, then computation time is reduced, but accuracy is reduced when sound sources are not on a grid
Solution Approach 1:
The patent transitions from grid-based discrete positioning to continuous spatial positioning by using coordinate-based target map generation. The sound source positions are represented as continuous coordinates rather than discrete grid points, allowing the neural network to accurately localize sound sources at any position in the spatial continuum while maintaining fast computation through efficient target map synthesis
3Productivity
If the maximum number of sound sources is predetermined in grid-free methods, then computation time is reduced, but the method cannot handle more sound sources than predetermined
Solution Approach 1:
The patent creates a universal target map generation mechanism that can handle any number of sound sources through coordinate-based positioning. The system synthesizes target maps by summing individual source contributions at arbitrary coordinates, allowing the same neural network architecture and training procedure to work for any number of sound sources without requiring predetermined limits or multiple specialized models
4Reliability
If beamforming methods are used, then robustness against measurement noise is improved, but spatial resolution becomes very low
Solution Approach 1:
The patent introduces a neural network as an intermediary between the beamforming map and the final sound source parameters. The beamforming map serves as a robust noise-resistant input, while the neural network acts as a mediator that learns to extract precise sound source positions and strengths from this input, effectively combining the noise robustness of beamforming with the high resolution capabilities of deconvolution methods
Data Source
Figure 1
Figure 2
Figure 3(a)~3(b)
AI summary
The present disclosure relates to a method and system for generating a target map as means for training a neural network and obtaining a sound source map regardless of the maximum number of sound sources, having a short computation time for inference, high spatial resolution and high sound source accuracy, and the method includes a grid generation step of generating grids each having a spacing within a given range at positions where sound sources are present in order to form a sound source map, a calculation step of calculating a result value for each of coordinates of the grids so that the result value is a local maximum at the position of a sound source and the result value decreases depending on the distance from the sound source, a matrix generation step of arranging the result values at positions on matrices corresponding to the respective coordinates of the grids, and a target map generation step of generating a target map having an image form by using the result values arranged in on the matrices.