Full-dimensional dynamic convolution and progressive learning strategy for strawberry recognition based on YOLOv8

Abstract

The growth of strawberries is influenced by environmental diversity and spatial dispersion, which present significant challenges for accurate identification and real-time image processing in complex environments. This paper addresses these challenges by proposing an advanced recognition model based on YOLOv8, tailored for strawberry identification. In this study, we enhanced the YOLOv8 architecture by replacing the traditional backbone with an EfficientNetV2 feature extraction network and using ODConv instead of the standard convolution. The loss function was modified with a dynamic nonmonotonic focusing mechanism, and WiseIoU was introduced to replace the traditional CIoU. The experimental results showed that the proposed model outperformed the original YOLOv8 regarding mAP50, precision, and recall, with improvements of 16.91%, 14.92%, and 8.4%, respectively. Additionally, the model's lightness increased by 15.67%. The proposed model demonstrated superior accuracy in identifying strawberries of different ripeness levels. The improvements in the proposed model indicate its effectiveness in strawberry recognition tasks, providing more accurate results in varying environmental conditions. The lightweight nature of the model makes it suitable for deployment on picking robots, enhancing its practical applicability for real-time processing in agricultural settings.