Domain-Aware Lightweight Spectral-Grouped Convolutions for Hyperspectral Fish Freshness Classification
SGNet, a lightweight architecture using grouped convolutions and a depthwise spatial pathway with dual attention, achieves 97.8% classification accuracy and 0.64 days MAE on a 16-day refrigerator-stored salmon fillet dataset with 4.75M parameters.
A research paper proposes SGNet for hyperspectral fish freshness classification. It separates spectral and spatial feature extraction using grouped convolutions and a depthwise spatial pathway, with dual attention combining channel-wise squeeze-and-excitation and spatial gating. On a newly developed 16-day refrigerator-stored salmon fillet dataset, SGNet achieves 97.8% accuracy and 0.64 days MAE with 4.75M parameters, a five- to eighteen-fold parameter reduction compared to ResNet-50 and Vision Transformers. Ablation studies validate each component.
The architecture exploits HSI-specific properties: spectral dominance over spatial textures, ordinal label structure, and small training samples. Grouped convolutions reduce parameters while preserving spectral bands; depthwise spatial pathway captures limited spatial cues; dual attention adaptively weights informative features. This domain-aware design enables high accuracy with low computational cost.
The approach targets industrial food quality assessment, offering nondestructive, real-time freshness prediction. Lightweight models are suitable for edge deployment in processing facilities. The dataset and method may encourage adoption of HSI-based automated inspection in seafood supply chains.
Potential to reduce manual inspection labor, improve freshness consistency, and minimize waste in seafood processing. The low parameter count suggests lower hardware requirements, enabling affordable deployment.
Next signals include validation on other fish species or storage conditions, integration into production inspection systems, and potential extension to other hyperspectral classification tasks. Commercialization may follow if robustness and cost-effectiveness are demonstrated.