Strawberry Inflorescence Detection Using Improved YOLOv5

To address the challenge of insufficient inflorescence detection accuracy during critical floral and fruit regulation phases in strawberry cultivation, this research presents an innovative detection framework based on a modified YOLOv5 architecture, designated YOLOv5-LCG. First, a$160 \times 160$large-scale target detection layer at the network's neck improves detection of small objects. Second, A lightweight upsampling operator CARAFE is utilized to collect more detailed information about the inflorescences because of the large-field aggregation capabilities, which allows it to acquire contextual information in a greater field of view. Finally, a GFPN structure is used to realize the different scale features between the interaction and fusion to improve the detection performance. The results of the experiment indicate that the modified model acquired an accuracy of 79.4 %, recall of 76.3 %, mAP of 78.1 %, and$\mathbf{F 1}$-score of$\mathbf{7 7. 8 \%}$for strawberry inflorescence recognition. Compared to the original model, this represents improvements of 4 percentage points of precision, 3.1 percentage points of recall, 4.1 percentage points of mAP, and 3.5 percentage points of the$F 1$score. YOLOv5-LCG demonstrates superior performance in strawberry inflorescence recognition, providing strong technical support for the flower and fruit thinning process in automated strawberry cultivation.