A cross-modal knowledge distillation framework transferring hyperspectral insights to smartphone edge computing for Agaricus bisporus freshness monitoring

Guo, Z., Wang, Y., Cao, L., Feng, L., Li, W., Xing, K., Auat Cheein, F. and Guo, X. (2026) A cross-modal knowledge distillation framework transferring hyperspectral insights to smartphone edge computing for Agaricus bisporus freshness monitoring. Smart Agricultural Technology, 15. ISSN 27723755

Full text not available from this repository.

Abstract

The real-time, non-destructive freshness monitoring of highly perishable Agaricus bisporus was identified as a critical task for postharvest supply chains. While hyperspectral imaging (HSI) demonstrated strong capabilities in capturing physiologically plausible quality assessments, its practical deployment was frequently hindered by high hardware costs and environmental constraints. To bridge the gap between advanced spectral analysis and accessible edge devices, an asymmetric cross-modal knowledge distillation (CMKD) framework was proposed. Initially, a hyperspectral teacher network, integrating a differentiable spectral gate-compression (DSGC) module and a bidirectional selective state-space (Mamba) encoder, was designed to extract physiological deterioration patterns from paired HSI-RGB observations. Subsequently, the learned spectral representations were distilled into a highly efficient RGB-only student network (LiteMobileStudent, 0.2 M parameters) via an annealed feature alignment and delayed logit-matching strategy. Operating exclusively on standard digital RGB images, the distilled student network achieved a classification accuracy of 81.11% and a regression root-mean-square error of 1.37 days. This performance effectively outperformed several standard lightweight convolutional architectures while maintaining a low computational footprint. To evaluate its practical feasibility, the framework was integrated into a smartphone-based edge computing application. Supported by a robust four-level hierarchical segmentation pipeline to mitigate complex background interferences in retail environments, the edge engine achieved an inference latency of 18.6 ms on a standard mobile CPU. By transferring spectrally-informed features into a mobile RGB interface, this study presented a scalable and computationally efficient AI solution, demonstrating practical potential for postharvest quality management in agricultural Internet of Things ecosystems

Item Type:	Article
Keywords:	White button mushroom, Teacher-student architecture, State-space models, Postharvest shelf life, Internet of things, Non-destructive evaluation
Divisions:	Departments > Engineering Research Centres > Harper Institute of Technology
Depositing User:	Miss Anna Cope
Date Deposited:	25 Sep 2026 09:17
Last Modified:	25 Sep 2026 09:17
URI:	https://hau.repository.guildhe.ac.uk/id/eprint/18429

Actions (login required)

Edit Item