WTSynNet: a lightweight cooperative network for multi-species Raman spectral classification

Abstract

Animal blood and semen contain diverse biochemical constituents that are of great importance in forensic science, veterinary diagnostics, and species traceability. Raman spectroscopy has emerged as a powerful tool for body fluid identification owing to its non-destructive and rapid acquisition of molecular vibrational fingerprints. However, achieving a balance between discriminative feature extraction and computational efficiency remains a challenge, particularly in imbalanced multiclass scenarios. To address this issue, we propose WTSynNet, a lightweight framework that integrates a one-dimensional wavelet convolution module (WTConv1d) with a star operation mechanism to enable efficient multiscale feature learning. Experiments on animal blood and semen Raman spectral datasets demonstrate that WTSynNet attains over 98% classification accuracy with fewer than 0.3 M parameters, while maintaining extremely low inference latency and memory usage. Moreover, the model achieves strong performance on a cross-domain marine pathogen Raman dataset, underscoring its robustness and adaptability. These results indicate that WTSynNet is a compact yet powerful model with strong generalization capability and holds broad potential for future applications in rapid on-site Raman spectral analysis.