EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Lightweight convolutional neural network for blueberry freshness classification and quality assessment

Tao Lu, Yan Xin, Kunhua Liu, Ting Liu, Xiaoyu Zhou, Yanting Yu, Yingying Wang

Journal of Food Composition and Analysis · 2026-11 · 159 · 109525

DOI: 10.1016/j.jfca.2026.109525 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Food Composition and Analysis, 2026-11. Verfasst von Tao Lu und 6 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: Natural Science Foundation of Qingdao Municipality, Natural Science Foundation of Shandong Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The freshness and quality of blueberries directly affect their nutritional value, commercial grading, and consumer acceptance. Traditional inspection methods are subjective, inefficient, and unsuitable for high-throughput applications, whereas many deep learning-based methods involve considerable computational cost, restricting their deployment on mobile and edge devices. To address these challenges, this study proposes DHAF-Net, a lightweight Dynamic Hierarchical Attention Fusion Network tailored to fine-grained blueberry freshness classification. Its methodological contribution lies in a coordinated hierarchical feature-learning framework rather than the independent use of existing attention and convolution operators. Specifically, ParallelSECA adaptively fuses channel recalibration and coordinate-aware spatial attention; the Dilated Convolution Sequence Optimizer refines degradation features using multiple effective receptive fields; and CrossDimDCMHAttention dynamically models interactions among low-, mid-, and high-level representations. Built on a pretrained MobileNetV3-Small backbone, DHAF-Net classified blueberries into nine storage-duration-based freshness levels and achieved 97.36% accuracy on a held-out image-level test set. Furthermore, quadratic polynomial regression models were established between continuous storage time (days) and the measured physicochemical attributes, including weight, firmness, and sweetness. By propagating the freshness classes predicted by DHAF-Net through these regression models, the physicochemical estimation errors were further quantified using mean absolute error (MAE) and root mean square error (RMSE), establishing a quantitative classification-to-quality-assessment pipeline.

Förderung

  • Natural Science Foundation of Qingdao Municipality
  • Natural Science Foundation of Shandong Province

Zitieren

APA

Lu, T., Xin, Y., Liu, K., Liu, T., Zhou, X., Yu, Y., & Wang, Y. (2026). Lightweight convolutional neural network for blueberry freshness classification and quality assessment. Journal of Food Composition and Analysis, 159, 109525. https://doi.org/10.1016/j.jfca.2026.109525

Vancouver

Lu T, Xin Y, Liu K, Liu T, Zhou X, Yu Y, et al. Lightweight convolutional neural network for blueberry freshness classification and quality assessment. Journal of Food Composition and Analysis. 2026;159:109525. doi:10.1016/j.jfca.2026.109525

BibTeX
@article{Lu2026,
  author = {Lu, Tao and Xin, Yan and Liu, Kunhua and Liu, Ting and Zhou, Xiaoyu and Yu, Yanting and Wang, Yingying},
  title = {Lightweight convolutional neural network for blueberry freshness classification and quality assessment},
  journal = {Journal of Food Composition and Analysis},
  year = {2026},
  volume = {159},
  pages = {109525},
  doi = {10.1016/j.jfca.2026.109525}
}
Kostenlos starten →