YOLOv11-GSF: an optimized deep learning model for strawberry ripeness detection in agriculture

The challenge of efficiently detecting ripe and unripe strawberries in complex environments like greenhouses, marked by dense clusters of strawberries, frequent occlusions, overlaps, and fluctuating lighting conditions, presents significant hurdles for existing detection methodologies. These methods often suffer from low efficiency, high computational expenses, and subpar accuracy in scenarios involving small and densely packed targets. To overcome these limitations, this paper introduces YOLOv11-GSF, a real-time strawberry ripeness detection algorithm based on YOLOv11, which incorporates several innovative features: a Ghost Convolution (GhostConv) convolution method for generating rich feature maps through lightweight linear transformations, thereby reducing computational overhead and enhancing resource utilization; a C3K2-SG module that combines self-moving point convolution (SMPConv) and convolutional gated linear units (CGLU) to better capture the local features of strawberry ripeness; and a F-PIoUv2 loss function inspired by Focaler IoU and PIoUv2, utilizing adaptive penalty factors and interval mapping to expedite model convergence and optimize ripeness classification. Experimental results demonstrate the superior performance of YOLOv11-GSF, achieving an average precision of 97.8%, an accuracy of 95.99%, and a recall rate of 93.62%, representing improvements of 1.8%, 1.3 percentage points, and 2.1% over the original YOLOv11, respectively. Furthermore, it exhibits higher recognition accuracy and robustness compared to alternative algorithms, thus offering a practical and efficient solution for deploying strawberry ripeness detection systems.