A Comparative Analysis of Deep Learning Architectures for Robust Fruit Detection

Accurate fruit detection is pivotal for advancing agricultural automation and precision farming. While convolutional neural networks (CNNs) have dominated this domain, their reliance on local features limits performance in cluttered environments. This study systematically evaluates transformer-based models against CNN baselines for fruit detection using the Fruits-360 dataset, comprising 110,297 images across 165 fruit categories. We tested 10 architectures, including CNNs (ResNet-18, DenseNet-121), pure Vision transformers (ViT-Small, GCViT-Xtiny), and hybrid models (MobileViTv2-050, EfficientViT-B0). Results demonstrated that transformer-based models, particularly GCViT-Xtiny, achieve state-of-the-art performance with 99.78% test accuracy, surpassing top CNNs like DenseNet-121 (99.30%). Hybrid models such as EfficientViT-B0 also excel (99.39% accuracy), balancing efficiency and global contextual understanding. However, pure transformer-based models like ViT-Small underperform (96.02%), highlighting data efficiency challenges. These findings underscore the potential of transformer-based models to overcome CNN limitations in fruit detection, though computational costs remain a consideration. Our work provides actionable insights for deploying robust vision systems in agricultural automation.