Recognition of Wood-Boring Insect Creeping Signals Based on Residual Denoising Vision Network

Area of Science:

• Agricultural Entomology
• Biotechnology
• Computer Science

Background:

• Current timber customs inspection for wood-boring pests relies on inefficient manual visual methods.
• Manual inspection suffers from long detection times, high labor costs, and reliance on human experience.
• Existing methods struggle to meet the demands of efficient and intelligent customs quarantine.

Purpose of the Study:

• To develop a rapid identification system for wood-boring pests using their peristaltic signals.
• To improve the efficiency and accuracy of customs quarantine for timber.
• To leverage deep learning for automated pest detection.

Main Methods:

• A hardware-software system (LabVIEW) collected pest peristaltic signals.
• Signals were processed, converted to audio, and features extracted using MFCC, PNCC, and RASTA-PLP.
• A Residual Denoising Vision Network (RDVNet) with multi-attention mechanisms was developed and trained.
• The denoising module (de-RDVNet) and classification performance were compared against established models.

Main Results:

• Power-Normalized Cepstral Coefficients (PNCC) demonstrated comprehensive feature extraction.
• The de-RDVNet achieved superior denoising performance with PNCC input (PSNR: 29.8, SSIM: 0.820).
• The RDVNet model achieved high classification accuracy (92.8%) and F1 score (0.878).

Conclusions:

• The developed peristaltic signal identification system effectively improves detection efficiency for wood-boring pests in timber.
• The RDVNet model shows significant potential for intelligent customs quarantine applications.
• This system offers practical value by reducing labor costs and enhancing detection speed.