Lightweight Deep Learning Framework for Fruit Freshness Classification with Knowledge Distillation and Grad-CAM Visualization
International Journal of Innovative Science and Research Technology · 09.04.2026 · 3670
DOI: 10.38124/ijisrt/26mar1828 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Innovative Science and Research Technology, online erschienen am 09.04.2026. Verfasst von Utsha Sarker und 5 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben.
Abstract (Originalsprache)
Fruit freshness evaluation has become an important task in food quality safety and food supply chain management, and traditional manual inspection method is subjective, time-consuming and inconsistent. Recent developments in deep learning have opened the door to automated quality analysis of fruit, but most of the top-performing models tend to be computationally expensive and hard to interpret and thus limit their use in the real world. In our work, we conduct a novel and efficient fruit freshness detection framework, which combines a high-capacity teacher network (ResNet-based) with an easy-to-train student model with knowledge distillation. To improve transparency, Gradient-weighted Class Activation Mapping (Grad-CAM) is used for the visualisation of discriminative regions that affect the model's prediction to increase trust and interpretability in the practical use Experimental results show that distilled student model achieves similar results compared to the teacher network (e.g. accuracy and F1-score results are within 1-3% margin) but with massive model size and inference latency reduction compared to the teacher network, which are consistent with the results obtained in recent studies on fruit classification under computing efficiency conditions [1], [2] and efficient knowledge distillation approaches. In addition, Grad-CAM visualisations also bring up relevant freshness indicators such as discoloration and texture variations, which are consistent with research on explainable AI-based fruit quality [3]. The framework proposed here offers an approach involving a good balance among accuracy, efficiency, and interpretability, which is suitable to be used in real world deployment applications in smart agriculture and food monitoring systems.
Zitieren
APA
Sarker, U., Vaishnav, L., Biswas, A., ., H., Ali, I., & Agarwal, P. (2026). Lightweight Deep Learning Framework for Fruit Freshness Classification with Knowledge Distillation and Grad-CAM Visualization. International Journal of Innovative Science and Research Technology, 3670. https://doi.org/10.38124/ijisrt/26mar1828
Vancouver
Sarker U, Vaishnav L, Biswas A, . H, Ali I, Agarwal P. Lightweight Deep Learning Framework for Fruit Freshness Classification with Knowledge Distillation and Grad-CAM Visualization. International Journal of Innovative Science and Research Technology. 2026:3670. doi:10.38124/ijisrt/26mar1828
BibTeX
@article{Sarker2026,
author = {Sarker, Utsha and Vaishnav, Lalit and Biswas, Archy and ., Harsh and Ali, Ikram and Agarwal, Priyanshu},
title = {Lightweight Deep Learning Framework for Fruit Freshness Classification with Knowledge Distillation and Grad-CAM Visualization},
journal = {International Journal of Innovative Science and Research Technology},
year = {2026},
pages = {3670},
doi = {10.38124/ijisrt/26mar1828}
}
EUVIMED