CGS-YOLOv11s: target detection algorithm for Korla fragrant pears in multiple natural environments

Abstract The color of Korla fragrant pears often closely resembles that of the surrounding leaves. Furthermore, mutual occlusion among fruits and foliage increases detection difficulty, making them harder to identify than common fruits under natural environments. To address this issue, a novel detection model, CGS-YOLOv11s, is proposed in this study to improve the accuracy of Korla fragrant pear detection under various natural environmental conditions. First, the C3CGK1 module is introduced to enhance feature extraction capabilities while reducing parameter count and interrupting redundant gradient propagation. Additionally, GhostConv is integrated to reduce the computational overhead of redundant features, further minimizing the model size and parameter volume. Furthermore, The Scylla Intersection over Union loss function (SIoU) is employed to enhance the precision of bounding box regression. In the experiments, CGS-YOLOv11s achieved a precision (P) of 91.6%, recall (R) of 84.1%, [email protected] of 92.1%, and [email protected]:0.95 of 72.0%, reflecting respective gains of 1.5%, 3.0%, 3.3%, and 3.2% over the baseline model. Furthermore, the model size was reduced by 34.0%, resulting in a compact model of only 12.0 MB. Compared with Faster R-CNN, SSD, YOLOv5s, YOLOv7-tiny, YOLOv8s, and YOLOv10s, CGS-YOLOv11s achieved [email protected] gains of 21.2%, 11.5%, 4.5%, 1.8%, 3.6%, and 4.2%, respectively. Through comparisons with other existing object detection models, CGS-YOLOv11s achieves a favorable balance between accuracy and model compactness. It can accurately identify targets even when interfered by leaves visually similar to Korla fragrant pears and performs well in densely distributed fruit scenarios. Moreover, the model maintains strong detection performance even with a relatively small number of small-sized pear samples. The research results enable real-time and precise identification of Korla fragrant pears, providing strong technical support for their automated harvesting in orchard environments under multiple natural environments.