Target detecting phalaenopsis seedlings using lightweight YOLOv8n-Aerolite

This study aimed to improve the efficiency of visual detection for the seedling gripping points in the automated rapid propagation of Phalaenopsis orchids, particularly on edge devices with limited computational resources and storage capacities. A lightweight algorithm of object detection was introduced (named YOLOv8n-Aerolite), in order to balance the high detection accuracy and the low computational complexity. As such, the algorithm was suitable for real-time applications on devices with restricted hardware capabilities. The StarNet was then developed as the backbone network, due to its efficient extraction of various features. An SPPF_LSKA (Large-Separable-Kernel-Attention) layer was incorporated to further optimize the model. The computational demands of the model were significantly reduced to maintain high precision during detection. The large-separable-kernel was designed to enhance the performance of the model, in order to process the key visual features with minimal resource usage. There was a critical advancement for the edge devices. Additionally, a new C2f_Star module was implemented to combine with the StarBlock in the network's neck for better feature fusion. Some finer details were then detected, such as the small and intricate points of seedling gripping. The C2f_Star was also integrated to introduce multi-scale feature processing. The gripping points were distinguished in the dense environments, where the seedlings were closely spaced. The detection head was also redesigned to include a lightweight shared convolutional layer structure, referred to as Detect_LSCD (Lightweight Shared Convolutional Detection Head). There was a notable increase in the detection speed, in order to reduce the overall size of the model. Specifically, the optimizations were fully realized to perform efficiently under resource-limited environments. The improved YOLOv8n-Aerolite algorithm was then tested on the image dataset of Phalaenopsis seedling. Experimental results showed that the improved model was achieved with an average inference speed of 435.8 frames per second, highly suitable for real-time applications. The improved model was also marked as one of the greatest available options for the detection of edge-based seedlings than before. The detection accuracy of the improved model reached 91.1%, particularly with an impressive precision of 91.6% to detect the small targets black tuber. The gripping points of the seedlings also validated the reliability of the improved model in practical deployments. Such high accuracy was achieved in the detection of small targets, indicating better suitability for tasks where the precise targeting of small objects was essential. In addition, the weight file size of the improved model was compressed to just 3.1 MB, particularly for the deployment on the edge devices where the storage capacity was constrained. A series of practical gripping experiments were conducted to further validate the algorithm. A success rate of 78% was obtained for the high efficiency of the improved model in real-world scenarios. The generalizability of the YOLOv8n-Aerolite algorithm was also tested on the 3D reconstructed dataset of phalaenopsis seedlings, similar to the detection of small targets. The results showed that the mAP_0.5 increased by 1.6 percentage points compared to the original YOLOv8n model. The better performance of the improved model was obtained across different datasets. The cross-dataset testing confirmed that the robustness and adaptability were suitable for a variety of detection tasks. In conclusion, the YOLOv8n-Aerolite algorithm significantly advanced the field of automated crop propagation. A highly efficient, accurate, and lightweight solution was also provided for the visual detection. The finding can serve as a valuable reference to develop scalable and automated technologies, especially for small-scale crops like Phalaenopsis orchids. YOLOv8n-Aerolite can also fully meet the needs of edge computing environments, particularly for the broader applications of agricultural automation.