Automated Tomato Quality Classification Using Transfer Learning And Machine Learning Techniques

Abstract

The increasing demand for high-quality tomatoes that satisfy consumer and market requirements, together
with large-scale agricultural production, has highlighted the need for automated inline quality grading
systems. Traditional manual grading methods are labor-intensive, time-consuming, and costly, making them
unsuitable for large-scale operations. This study proposes a hybrid approach for tomato quality sorting and
grading by combining pre-trained convolutional neural networks (CNNs) for feature extraction with
conventional machine learning algorithms for classification. A tomato image dataset was created using an
NVIDIA Jetson TX1 single-board computer, followed by image preprocessing and fine-tuning techniques to
enable deep networks to capture complex and discriminative features. The extracted features were then
classified using traditional machine learning classifiers, including Support Vector Machine (SVM), Random
Forest (RF), and K-Nearest Neighbors (KNN). Experimental results show that the proposed CNN–SVM
hybrid model outperforms other approaches, achieving an accuracy of 96.2% for binary classification (healthy
vs. rejected) and 95.4% for multiclass classification (ripe, unripe, or rejected) when InceptionV3 is used as the
feature extractor. Furthermore, evaluation on a public dataset demonstrated that the CNN–SVM model
achieved an accuracy of 96.8% in classifying tomatoes into ripe, unripe, old, and damaged categories,
surpassing other hybrid models. The performance of the proposed system was comprehensively assessed
using metrics such as accuracy, precision, recall, specificity, and F1-score, confirming its effectiveness for
automated tomato quality grading.