Year: 2026
Type: article
Abstract: Economically motivated adulteration (EMA) of blueberry puree poses severe food safety and market risks. Conventional single-modality non-destructive testing struggles to simultaneously capture the macroscopic morphology and microscopic chemical attributes of heterogeneous solid-liquid mixtures, while simple multimodal concatenation remains vulnerable to industrial noise. To overcome these bottlenecks, this study proposes CompleteNet, a deep multimodal fusion network driven by bidirectional cross-attention for robust, online adulteration detection. CompleteNet leverages a cascaded 1D-CNN to extract near-infrared (NIR) overtone features and a dual-attention 3D-CNN to disentangle hyperspectral imaging (HSI) spatial-spectral features. Crucially, a multi-scale spectral attention mechanism dynamically calibrates critical wavebands entirely based on intrinsic physicochemical differences, significantly enhancing physical interpretability without relying on predefined empirical wavelengths. A latent space tokenization-based bidirectional cross-attention bridge actively aligns macroscopic contexts with microscopic features, substantially reducing the semantic gap. Furthermore, an asymmetric multi-task learning head simultaneously discriminates 13 adulterant categories and predicts concentrations (0–40%). Results demonstrated that CompleteNet achieved a peak classification accuracy of 99.76% and a prediction determination coefficient ( ) of 0.9879. Under simulated industrial noise, the architecture exhibited exceptional anti-interference and feature self-healing capabilities, vastly outperforming baseline models. Finally, deploying the optimized model on an edge-computing platform (NVIDIA Jetson Orin Nano) enabled real-time analysis at ∼15 frames per second (FPS). This study delivers a highly robust, interpretable edge-deployment framework for the high-throughput, in-line monitoring of complex food matrices.
Source: LWT
Institution: Northeast Forestry University
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $3,240
OpenAlex ID: W7204235072