TL-YOLOv8: A Blueberry Fruit Detection Algorithm Based on Improved YOLOv8 and Transfer Learning

Abstract

Blueberry fruit in the picking process are characterized by small fruit particles, similar color characteristics of immature fruits to leaf, and not obvious characteristics of fruits obscured by leaves, which lead to low real-time detection accuracy of blueberry fruits and fail to meet the detection standard of automatic picking. To improve the detection accuracy, we propose a blueberry identification algorithm named TL-YOLOv8 based on the YOLOv8 algorithm. By introducing an improved MPCA (Multiplexed Coordinated Attention) in the last layer of the backbone network, we enhance the feature extraction capability during the training process. By replacing the C2f module with an OREPA(Online Convolutional Re-parameterization) module, not only the training is accelerated, but also the characterization is enhanced. Meanwhile, in order to cope with the fruit occlusion problem more effectively, we introduce the MultiSEAM (Multi-scale Separation and Occlusion-Aware Module). As a means of optimizing the model parameters, we pre-trained the model using Transfer Learning to improve the generalization ability of the network. The method achieved 84.6% Precision, 91.3% recall and 94.1% mAP on the blueberry dataset, where mAP was improved by 3.4% over the original algorithm, and the experiments showed that it can be effective for blueberry fruit detection.

Keywords

• Blueberry detection
• deep learning
• object detection
• occlusion
• transfer learning
• YOLOv8