Multi-layer feature fusion and attention-enhanced YOLOv9 for rapid field detection of greenhouse blueberry maturity at long and short distances

Accurate maturity monitoring is critical for blueberry growth management and digitalization. Given the dense growth of blueberry plants and small fruit size, detecting blueberry maturity at long distances poses significant challenges. Therefore, this paper presents an improved algorithm, YOLOv9-MSAS, for detecting greenhouse blueberry maturity at long and short distances. YOLOv9-MSAS includes a multi-layer hierarchical weighted feature fusion module (MHWFuse), which adopts a layer-by-layer weighted upward fusion method to enhance the perception of feature information at a long distance. To increase the weights of relevant regions and useful features, a feature extraction module with spatial and channel attention mechanisms, RepNCSPELAN4-A, is designed. In addition, a novel upsampling and downsampling module and the ShapeIoU loss function are employed. The experimental results demonstrate that the mean average precision (mAP) of the YOLOv9-MSAS network reached 85.8%, which was 4.6% higher than that of the YOLOv9 model, and that the FPS reached 56.4. Compared with that of YOLOv9, the mAP of YOLOv9-MSAS was improved by 2.9% and 3.7% at short and long distances respectively. Ablation experiments and GradCAM heatmaps were used to verify the effectiveness of each module of YOLOv9-MSAS. Finally, the generalization ability and robustness of YOLOv9-MSAS were validated on an open-source strawberry dataset.