CHE-YOLOv8: A Lightweight Model for Front Morphology Detection of Pepper Flowers

The study focuses on the small-object detection task of the front morphology of pepper flowers, aiming to improve detection accuracy in complex shapes through lightweight design. To address this challenge, an improved lightweight model, CHEYOLOv8, based on the YOLOv8n architecture, is proposed. The model enhances the capture of long-range pixel correlations by optimizing the backbone network structure and incorporating the C2f-CAA-Star module. Additionally, the High-order Selection Feature Pyramid Network (HSFPN) is used to optimize multiscale feature extraction and fusion, significantly improving the model's ability to detect small targets in complex backgrounds. The detection head employs the Efficient Multi-scale Convolution (EMS-Conv) module, further enhancing detection accuracy and robustness. Experimental results demonstrate that the improved YOLOv8 model excels in both detection accuracy and efficiency: the $\text{mAP} {@} 0.5$ increases to $92.3 \%, \text{mAP} {@} 0.5: 0.95$ reaches 5 8. 4 %, while reducing computational complexity by 32.1 % and the number of parameters by 38.9 %.