FruitNet: a deep hybrid model for fruit detection and yield estimation

Abstract

In recent years, intelligent agriculture monitoring systems have attracted considerable interest for yield estimation and fruit quality inspection. This paper introduces FruitNet, a deep hybrid model integrating object detection, classification, and regression for automated apple detection and yield prediction. The architecture employs YOLOv8 for real-time detection, a convolutional neural network (CNN) for quality assessment across four categories (excellent, good, average, and bad), and random forest regression for estimating yield based on extracted features. To enhance classification robustness, the CNN features are further refined using a support vector machine (SVM) classifier, tuned via grid search for optimal performance. The system is implemented using Python and Django, with a preprocessing pipeline incorporating noise removal, data augmentation, and normalization. The model is trained and evaluated on the MinneApple dataset containing over 18,000 annotated images. Experimental results demonstrate high generalization performance with over 98% training accuracy and 94–95% validation accuracy across 15 epochs. Visual analytics including confusion matrices and detection overlays confirm robust detection and classification. The proposed FruitNet framework shows strong potential for deployment in real-time precision agriculture, supporting mobile integration, orchard-level insights, and scalable smart farming applications.