다중 자가집중 유-넷 스타일 인셉션 신경망 기반 초음파 카메라 센서의 신호 왜곡 복구

초록

Distortion in ultrasound signals occurs when a source with energy exceeding the reception capacity of an ultrasound device is input, making it difficult to recognize the original signal's information. Particularly, the ultrasonic frequency band (above 20 kHz) contains complex patterns due to its high-frequency characteristics. When it is, combined with a large number of samples, it makes distortion recovery more challenging than in audible or low-frequency bands. To address these challenges, this paper proposed a design leveraging an inception structure, applying convolutional filters of various sizes to simultaneously extract features across different time scales. This approach allowed the model to effectively capture characteristics of ultrasound signals from both microscopic and macroscopic perspectives. Multi-scale features were then compressed through an encoder, followed by a Multi-Head Self-Attention mechanism that could learn crucial correlations within complex ultrasonic signal patterns, identifying important features lost due to distortion. Finally, these features were reconstructed into high-resolution raw waveforms through a Wave U-Net architecture. The proposed method was rigorously evaluated using a discharge signal dataset recorded with an ultrasound camera sensor at various distances, demonstrating higher performance in MSE, cosine similarity, and SSIM metrics than other models, with a notable performance improvement of over +2 dB in SI-SDR.

키워드