ResearchPod Summary
Assessing fish freshness is critical for food safety and quality control, but traditional methods are often destructive or subjective. While hyperspectral imaging (HSI) provides rich biochemical data, existing deep learning models—typically designed for remote sensing or natural images—struggle with the unique constraints of food quality data: high spectral dimensionality, limited training samples, and an ordinal label structure (where freshness degrades over time).
The authors propose SGNet, a lightweight architecture specifically engineered for HSI food data. Instead of using a generic backbone, SGNet employs a hierarchical encoder that separates spectral and spatial processing. It uses grouped pointwise convolutions to manage spectral information without premature entanglement, and a depthwise spatial pathway to capture local tissue texture. A dual attention mechanism, which combines channel-wise squeeze-and-excitation with spatial gating, allows the model to highlight relevant features with minimal computational overhead. The model was trained and tested on a newly curated 16-day refrigerated salmon dataset, ensuring no data leakage by separating samples by fish pack.
SGNet achieves 97.8% classification accuracy and a mean absolute error (MAE) of 0.64 days, outperforming larger, standard architectures like ResNet-50 and Vision Transformers. The model requires only 4.75M parameters, representing a five- to eighteen-fold reduction in size compared to the baseline models tested. The results indicate that the model is highly effective at identifying freshness within the commercially relevant window, with errors primarily occurring only in the advanced stages of spoilage.
This research demonstrates that domain-specific inductive biases—such as separating spectral and spatial processing—are more effective than raw model capacity for HSI-based food quality assessment. By providing a lightweight, high-accuracy solution, SGNet offers a viable path for real-time, non-destructive freshness monitoring in industrial settings where computational resources may be limited.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.