Intelligent Detection of Fruit Decay Severity Using YOLO with Hybrid Attention Mechanisms

Abstract

Accurate detection of fruit decay severity is crucial in agricultural supply chains for minimizing economic losses, ensuring food safety, and optimizing storage–transportation workflows. Traditional detection methods relying on manual inspection or simple image processing suffer from low efficiency and poor robustness. This study proposes an improved YOLO (You Only Look Once)-based method for fruit decay severity detection. A tailored data augmentation strategy is employed to enhance model adaptability to complex scenarios. Furthermore, Squeeze-and-Excitation (SE) channel attention and Convolutional Block Attention Module (CBAM) hybrid attention modules are integrated into the YOLOv8 backbone to strengthen feature learning for decay regions. Experiments on public fruit datasets compare the performance of YOLOv5, YOLOv8, and their attention-enhanced variants (YOLOv8 + SE and YOLOv8 + CBAM). The results demonstrate that the YOLOv8 + CBAM model achieves 92.3% mAP@0.5, a 3.9% improvement over the baseline, with superior performance in small-target detection and scenarios with complex background interference. This study innovatively combines multi-scale feature fusion and attention mechanisms, providing an efficient and precise solution for agricultural automation detection.