METHOD FOR DETERMINING SPECTROGRAM PARAMETERS OF RADIO SIGNALS FOR THEIR RECOGNITION USING COMPUTER VISION METHODS

Automatic recognition of radio signals using convolutional neural networks that analyze their spectrograms is a promising approach. However, its effectiveness is hindered by the lack of standardized parameters for generating these spectrograms, which are necessary to ensure their informativeness. To address this issue, a methodology was developed to determine the most suitable parameters for generating spectrograms of shortwave (high frequency) signals in the Adobe Audition environment based on expert evaluations. This approach enables the creation of a standardized dataset for training convolutional neural networks. An experiment was conducted involving experienced signals intelligence operators who individually adjusted spectrogram visualization parameters (size of fast Fourier transform, window function type, dynamic range, color scheme, and segment duration) for two typical high frequency signal types to achieve maximum visual informativeness. The collected data were statistically analyzed (mode, mean) to derive generalized recommendations. A unified recommended set of spectrogram generation parameters (size of fast Fourier transform, window function type, color scheme, dynamic range, and segment duration) was identified, reflecting expert consensus on the optimal visual representation of the analyzed signals. The proposed methodology and resulting recommendations allow for standardization of the spectrogram preparation process for high frequency signals in tasks of automatic recognition using convolutional neural networks, contributing to the development of high-quality datasets and improved reproducibility of research.