MSAMixerNet: A multi-scale MLP-Mixer model with attention mechanisms for Raman spectral classification of animal...

With the rapid development of the animal leather market, achieving accurate identification of leather from different animal species has become a critical issue for ensuring market regulation and protecting consumer rights. Raman spectroscopy, renowned for its non-destructive, rapid, and highly sensitive nature, has been extensively applied in the qualitative and quantitative analysis of complex materials. However, practical applications often encounter challenges such as high-dimensional redundancy, noise interference, spectral peak shifts, and category overlap, hindering traditional models from achieving stable and reliable classification performance. To address these issues, this study proposes a calibration-driven multi-model consensus learning framework-MSAMixerNet. This framework integrates a multi-scale MLP-Mixer backbone network, a Transformer-based global modeling module, a CBAM channel-spatial attention mechanism, and a positional encoding mechanism, thereby balancing local sensitivity with global structural modeling capabilities. Concurrently, it introduces a Bagging multi-model ensemble strategy. By constructing multiple submodels on different training subsets and performing consensus voting, it effectively mitigates overfitting and data distribution imbalance issues. Experimental results on animal leather Raman spectroscopy classification demonstrate that this framework achieves an accuracy of 98.27% across nine animal leather datasets, outperforming existing mainstream methods in both robustness and confidence reliability. This research not only provides a novel solution for intelligent spectral recognition in high-dimensional noisy environments but also offers a feasible technical pathway for subsequent market regulation and material anti-counterfeiting applications.