Abstract:

Objective: The existing blueberry fruit ripeness detection methods exhibit poor performance in complex natural environments and insufficient robustness against lens defocus blur and multi-angle imaging in actual harvesting operations, resulting in high contamination rates of unripe fruits and difficulties in ensuring harvest quality. Aiming at the above issues, an improved YOLOv9 detection method is proposed to achieve high-precision ripeness recognition, providing algorithmic support for vision-based dynamic adjustment of harvesting speed. Method: Based on the YOLOv9 model, MobileNetV4 was introduced into the YOLOv9 model as the backbone feature extraction network to reduce the number of parameters and computational burden of the network. Additionally, a GAM attention module was integrated into the neck network of YOLOv9, and the weight of each feature was adjusted to enable the model to better focus on the most important feature areas for target detection, thereby enhancing its ability to recognize key regions and improving detection accuracy and robustness. WIoU was employed as the loss function to optimize the model's localization accuracy, improve boundary box prediction accuracy, and accelerate network convergence. Blueberry harvesting experiments were conducted on a picking test platform to verify whether the model meets the precision and speed requirements for blueberry harvesters, and to determine the optimal harvesting speed when harvesting blueberries with different fruit maturity ratios. Result: The improved YOLOv9 model achieved a precision of 98.0%, recall of 97.2%, an average precision mean (mAP) of 98.2%, and a detection frame rate of 86.5 fps on the test set. Compared with SSD, Faster R-CNN, YOLOv5, and YOLOv8 models, the average precision mean improved by 6.8, 5.6, 4.0, and 2.7 percentage points, respectively. The improved model met the requirements of the harvesting system. When the proportion of mature fruits on blueberry plants was 90%~100%, 85%~90%, and 80%~85%, the optimal harvesting speeds were 125 r·min^?1, 130 r·min^?1, and 140 r·min^?1, respectively. Conclusion: The improved YOLOv9 model has significantly enhanced detection performance compared to the original model. The optimal harvesting speeds obtained through blueberry harvesting tests can reduce the rate of unripe fruits, providing strong technical support for the intelligent harvesting of blueberries.

Key words: blueberry maturity detection, YOLOv9, MobileNetV4, GAM, WIoU