Complex Spectrogram Color Conversion for Deep Learning Signal Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for converting RGB-based image information into HSV color space fail to accurately process spectrogram data for signal detection or identification using deep learning due to differing degrees of freedom per pixel, leading to inefficiencies in signal detection or identification.
Innovation Solution
A signal processing device employing a window function unit, Fourier transform unit, color space conversion unit, and deep learning processing unit to convert phase angles of complex amplitudes into hues and associate absolute values with brightness, generating spectrogram data for accurate signal detection or identification using a two-dimensional time-frequency plane.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional RGB to HSV color space conversion is applied to spectrogram data, then the processing can be performed using existing deep learning frameworks, but the detection accuracy deteriorates due to mismatched degrees of freedom per pixel
Solution Approach 1:
The patent changes the parameter representation from standard RGB/HSV color spaces to a custom color space where complex numbers are mapped to RGB values. This involves transforming the magnitude and phase of complex spectrogram values into RGB components using specific mathematical relationships, thereby adapting the data format to match the expectations of deep learning models while preserving the essential information needed for accurate signal detection.
2Loss of information
If complex number information is directly used in deep learning processing, then the information symmetry is maintained, but the compatibility with conventional deep learning frameworks deteriorates
Solution Approach 1:
The patent introduces an intermediary conversion layer that transforms complex number representations into RGB color space. This intermediary step serves as a bridge between the complex number domain (which preserves information symmetry) and the RGB domain (which is compatible with conventional deep learning frameworks). The conversion maintains the essential information through mathematical relationships while enabling compatibility with existing model architectures.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A signal processing device (1) includes: a window function unit (11) that performs window function processing on a Fourier transform frame cut from time-series signal data; a Fourier transform unit (12) that transforms, into information in a frequency domain, the Fourier transform frame subjected to the window function processing; a color space conversion unit (13) that generates spectrogram data by converting a phase angle of a complex amplitude in the frequency domain into a hue and performing conversion that associates an absolute value of the complex amplitude with brightness by mapping, the spectrogram data being defined so as to correspond to a two-dimensional plane of time and frequency; and a deep learning processing unit (14) that detects or identifies a signal by convolutional processing and fully connected processing on the spectrogram data.