Fruit Quality Classification Through Feature Fusion and Deep Learning

Fruit quality assessment is an integral part of food safety, loss reduction after harvest and consumer satisfaction. In modern agricultural production systems, where quality control is essential for ensuring both domestic and export markets, it continues to play a vital role. Manual grading, which is widely used, is poorly time-efficient, imprecise, and prone to human error, and is reliant upon the human eye and human skill. Such limiting practices can result in variability and, in large-scale production, high operating expenses. In this paper, we propose a hybrid classification model using deep learning features extracted from ResNet50V2 along with traditional handcrafted features such as Histogram of Gradients (HOG), Local Binary Patterns (LBP), and color, texture and shape descriptors. The purpose of this study is to develop a model to combine high-level and low-level feature representation to gain a comprehensive understanding of fruit quality characteristics. The proposed preprocessing technique will be followed by data augmentation, ANOVA-based feature selection, and classification using K-Nearest Neighbors (KNN), Random Forest (RF) and Extreme Gradient Boosting (XGBoost). This multi-stage course of action is specifically designed so that the extracted features are appropriate, relevant and discriminative for correct classification. For experimentation, standard datasets were chosen as FruitNet and tomato. There are six fruit types included in the FruitNet dataset with three different quality classes, and another tomato dataset with three ripeness-based classes. XGBoost always outperforms other classifiers on the FruitNet dataset with an accuracy of 97.3% and on the tomato dataset with an accuracy of 92%. The findings validate the validity of the proposed model and its use in all crops under all grading systems.