Pepper-YOLO: an lightweight model for green pepper detection and picking point localization in complex environments

In the cultivation of green chili peppers, the similarity between the fruit and background color, along with severe occlusion between fruits and leaves, significantly reduces the efficiency of harvesting robots. While increasing model depth can enhance detection accuracy, complex models are often difficult to deploy on low-cost agricultural devices. This paper presents an improved lightweight Pepper-YOLO model based on YOLOv8n-Pose, designed for simultaneous detection of green chili peppers and picking points. The proposed model introduces a reversible dual pyramid structure with cross-layer connections to enhance high-and low-level feature extraction while preventing feature loss, ensuring seamless information transfer between layers. Additionally, RepNCSPELAN4 is utilized for feature fusion, improving multi-scale feature representation. Finally, the C2fCIB module replaces the CIB module to further optimize the detection and localization of large-scale pepper features. Experimental results indicate that Pepper-YOLO achieves an object detection accuracy of 82.2% and a harvesting point localization accuracy of 88.1% in complex scenes, with a Euclidean distance error of less than 12.58 pixels. Additionally, the model reduces the number of parameters by 38.3% and lowers complexity by 28.9%, resulting in a final model size of 4.3MB. Compared to state-of-the-art methods, our approach demonstrates better parameter efficiency. In summary, Pepper-YOLO exhibits high precision and real-time performance in complex environments, with a lightweight design that makes it well-suited for deployment on low-cost devices.