Dual-Branch Multiscale Attention Network (DMANet): Integrating Dynamic Two-Dimensional Correlation Spectroscopy Using a Dual-Branch Multiscale Attention Network for Beef Adulteration

Abstract

To address the detection of low-level duck meat adulteration in beef, this study proposes a method based on near-infrared (NIR) two-dimensional correlation spectroscopy (2D-COS) and a dual-branch multiscale attention network (DMANet), enabling semi-quantitative prediction of duck meat content. The approach expands one-dimensional (1D) NIR spectra into synchronous and asynchronous 2D-COS, constructing a dual-branch network that extracts complementary characteristics using multiscale convolution. An adaptive attention mechanism dynamically fuses the two spectral modalities, enhancing sensitivity to subtle adulteration traits. Experiments show the model achieves 100% overall accuracy on the test set, outperforming a traditional partial least squares (PLS) model (81.6%) and single-input models (94.6% for synchronous and 92.8% for asynchronous maps). In the 0.5–5% adulteration range, it improves accuracy by 31.3%, 12.5%, and 18.7% over these benchmarks, respectively. Notably, 2D-COS consistently surpasses 1D spectral models in detecting low-content adulteration by amplifying dynamic spectral responses and strengthening characteristics of trace contaminants.

Graphical abstract

This is a visual representation of the abstract.