Výskumný ústav potravinársky

Wang, Y. P. – Li, Y. S. – Chen, R. Z. – Zhang, X. S. – Polovka, M. – Tobolková, B.
A mango cold damage bruise detection model based on multiscale flexible sensing and multimodal fusion learning
Journal of Food and Nutrition Research, 64, 2025, č. 4, s. 295-308

Xiaoshuan Zhang, College of Engineering, China Agricultural University, 100083 Beijing, China; Sanya Research Institute of China Agricultural University, Sanya Yazhou Bay Science and Technology City, 572000 Sanya, China. E-mail: zhxshuan@cau.edu.cn

Original article
Received 9 August 2025; 1st revised 10 October 2025; accepted 21 October 2025; published online 7 November 2025.

DOI: https://doi.org/10.64122/GTDF1855

Súhrn: Traditional grading methods rely on experts’ subjective sensory evaluation to classify fruits based on their appearance, texture, and other characteristics. However, this approach suffers from significant subjectivity and inefficiency. To address these issues, this study proposes a multimodal classification decision system based on red-green-blue (RGB) images, thermal imaging (TIR), and impedance information (IMP) for the efficient detection of cold damage bruises in mangoes. Objective quality labels for mangoes were established using expert assessments. A multimodal enhanced convolutional neural network (MECN-net) was developed, integrating appearance features from RGB images, internal characteristics from TIR, and dielectric properties from IMP. MECN-Net enhances IMP feature extraction via transposed convolution and performs feature-level fusion with RGB and TIR data. The fused data is then classified using an improved AlexNet module, and decision vectors are further refined through fully connected layers to improve classification accuracy. The study also analysed the effects of different modality combinations on feature extraction. MECN-Net achieved a test accuracy of 95.5 %, precision of 98.7 %, recall of 98.4 %, and an F1 score of 0.985, outperforming existing models. These findings demonstrate the system’s potential as a reliable tool for automated fruit quality assessment and damage detection.

Kľúčové slová: mango classification; convolutional neural network; multimodal learning; feature fusion; decision system

Na stiahnutie:
  jfnr-2025-4-pp295-308-wang.pdf (PDF, 1.69 Mb, 124x)