VisDist-Net: A New Lightweight Model for Fruit Freshness Classification


Demirel S., YILDIZ O.

FOOD ANALYTICAL METHODS, 2024 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Publication Date: 2024
  • Doi Number: 10.1007/s12161-024-02716-4
  • Journal Name: FOOD ANALYTICAL METHODS
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Agricultural & Environmental Science Database, CAB Abstracts, Compendex, Food Science & Technology Abstracts, INSPEC, Veterinary Science Database
  • Gazi University Affiliated: Yes

Abstract

Agricultural production is of vital importance for humanity and the agricultural economy. Enhancing food security in agriculture can increase agricultural production and also help alleviate food scarcity. Also, the early detection of plant diseases can be crucial for quality agricultural products. The use of embedded software in Internet of Things devices for quality control processes has become quite widespread. These software applications require lightweight models. Therefore, a new model named the vision distillation network (VisDist-Net) has been developed to address real-world problems in agricultural production. This model aims to increase agricultural productivity by classifying three different fruits as rotten and fresh. An open-source dataset was used for this classification. VisDist-Net is a model created based on knowledge distillation. In the VisDist-Net model, knowledge is distilled from a vision transformer to a hybrid convolutional neural network (cnn). The strengths of both models have been combined by creating a hybrid student convolutional neural network through the fusion of feature vectors from resnet18 and mobilenetv1 models. This distillation process enables the creation of a high-performance lightweight model suitable for real-world applications. The VisDist-Net model has yielded quite promising results in this endeavor, achieving an f1-score of 0.9945 and an area under the curve (AUC) score of 0.9967.