A lightweight and optimized deep learning model for detecting banana bunches and stalks in autonomous harvesting vehicles

• The proposed novel architecture significantly enhances the overall performance of the detection model. • The optimized model achieves superior performance, with precision, recall, and mAP50 metrics of 96.3%, 90%, and 94.5%, respectively, surpassing the baseline and other well-known models. • This enhanced design optimizes the model’s parameters (1.7M) and size (3.7MB), reducing complexity by over 40% compared to the baseline model and enabling efficient deployment on embedded systems. • The proposed detection framework has been successfully deployed on a banana harvesting vehicle in the field, reducing labor costs and increasing productivity in the banana harvesting. . Developing algorithms to identify fruit cutting locations is important for the functionality of harvesting robots. However, existing studies often rely on multi-stage detection processes. This complicates system design and hinders real-time performance. To address these challenges, this study proposes a novel detection model for banana-harvesting robots. The model simultaneously detects banana bunches and stalks in orchard environments. It is built upon the YOLOv8n (You Only Look Once version 8 nano) baseline and includes enhancements to improve accuracy while preserving a lightweight architecture. Specifically, the standard convolution layers are upgraded with a lightweight group-shuffle convolution module, reducing complexity while preserving efficiency. Additionally, a novel C2f-fast efficient channel attention module is proposed in the backbone, significantly enhancing the model's feature extraction capabilities. Furthermore, the bidirectional feature pyramid network is introduced in the original neck network, improving feature aggregation and adaptability to varying environmental conditions. Experimental results demonstrate that the proposed model achieves performance, with precision, recall, and mAP50 metrics of 96.3%, 90%, and 94.5%, respectively, exceeding the baseline model by 0.5%, 2.6%, and 1%. Moreover, the parameters and size of the proposed model are optimized to 1.7M and 3.7MB, reflecting reductions of 43.3% and 40.3%, respectively, in comparison to the baseline. Notably, the proposed model outperforms the previous detection models, offering high accuracy while optimizing computational efficiency. These advancements make the proposed model highly suitable for deployment on embedded systems in agricultural robots.