SGNet: Lightweight Spectral-Grouped Network for Hyperspectral Fish Freshness Classification
A new lightweight deep learning framework named SGNet (Spectral-Grouped Network) has been developed by researchers to assess fish freshness through hyperspectral imaging (HSI). This innovative approach tackles significant issues in HSI data, such as the predominance of spectral information over spatial textures, the structure of ordinal labels, and a scarcity of training samples. SGNet employs grouped convolutions and a depthwise spatial pathway to differentiate between spectral and spatial feature extraction, further improved by a dual attention mechanism that integrates channel-wise squeeze-and-excitation with spatial gating. Evaluated on a novel dataset of salmon fillets stored in a refrigerator for 16 days, SGNet achieved a remarkable 97.8% accuracy and a mean absolute error (MAE) of 0.64 days, utilizing just 4.75 million parameters. The findings are published in a paper on arXiv (arXiv:2608.12227) and suggest potential for real-time, nondestructive freshness evaluations in the seafood sector, minimizing waste and enhancing quality.
Key facts
- SGNet is a lightweight architecture for hyperspectral fish freshness classification.
- It uses grouped convolutions and a depthwise spatial pathway to separate spectral and spatial features.
- A dual attention mechanism couples channel-wise squeeze-and-excitation with spatial gating.
- SGNet achieved 97.8% classification accuracy and 0.64 days MAE.
- The model has only 4.75 million parameters.
- Tested on a newly developed dataset of 16-day refrigerator-stored salmon fillets.
- Ablation studies validated the contribution of each component.
- The paper is available on arXiv with ID 2608.12227.
Entities
—