Design and experiment of the performance detection system for leafy vegetable plug seedling planting based on YOLO11n

High single-seeding rate and low missed-seeding rate are often required in the process of plug seedling seeding for leafy vegetables. Alternatively, the architecture and substantial parameters of YOLO11 can represent an advanced image detection framework. The remarkable accuracy can be expected to detect the diverse targets in the high-resolution and complex scenes. However, the high precision of the YOLO11 is also confined to the serious computational demands and long inference times, thus restricting its practical deployment in resource-constrained scenarios. In contrast, the lightweight YOLO11n model can be expected to reduce the computational complexity and parameter redundancy for the competitive performance after architectural optimizations. This study aims to propose an improved lightweight model (named Seed-YOLO) using You Only Look Once 11 nano (YOLO11n). The seeding performance was also detected for the three types of leafy vegetable seeds in the plug seedling trays. The model was then deployed on the edge computing device (NVIDIA Jetson Xavier NX). An efficient detection system was developed for the high-performance plug seedling seeding. Four components were utilized to improve the Seed-YOLO model. 1) A Context Anchor Attention (CAA) module was introduced into the backbone network to construct the C2PSA_CAA module. The feature representation of the seed center region was precisely enhanced to capture seed characteristics. The CAA module was a specific network structure to capture the long-range contextual information. Statistical features of the local regions were extracted after average pooling operations, and then strengthened using 1×1 convolutions, thereby enhancing the feature representation of seed central areas. The horizontal (1×11) and vertical (11×1) depth-wise separable strip convolutions were adopted to expand the receptive field for efficient computational complexity, similar to large convolution kernels. An attention weight map was generated via a Sigmoid function. The weight was then applied to the original feature map to realize weighted enhancement of the features. 2) Group Shuffle Convolution (GSConv) and GSBottleneck modules were incorporated into the neck network. The C3K2_GS module was then constructed to accelerate the fusion of seed features for the detection accuracy. Among them, the GSConv was a lightweight convolution. After the Shuffle operation, the feature information generated by standard convolution was evenly spread into every part by Depthwise Separable Convolution (DSC) over different channels. The computational complexity and the number of parameters were reduced to maintain the performance. 3) Wise Intersection over Union version 3 (WIoU v3) loss function was adopted to effectively anchor the boxes of average-quality seeds. Thereby, its dynamic non-monotonic focusing mechanism was employed to improve the detection performance. WIoU v3 was often used as a bounding box loss function. The gradient gains of samples were dynamically adjusted with different qualities using a weight factor. While reducing the focus on the high-quality samples, the negative gradients generated by low-quality samples were also mitigated to enhance the overall performance of the model. 4) An XSmall detection head was added to boost the detection accuracy for the small targets of the leafy vegetable seeds. While the original Medium/Large detection heads were removed to reduce the parameter count and size, thus achieving model lightweighting. Experimental results demonstrate that the Seed-YOLO was achieved in a mean average precision at 50% IoU (mAP@0.5) of 96.7% and F1 of 93.79% for the seeding performance of the three leafy vegetable seeds, indicating the improvements of 5.4 and 8.87 percentage points, compared with the YOLO11n’s 91.3% and 84.92%, respectively. Notably, the model’s parameter count was reduced to 1.58 million, which was a 38.7% decrease from YOLO11n’s 2.58 million. The model was deployed on the NVIDIA Jetson platform. A graphical user interface was developed in a real-time detection system for the plug seedling seeding. When operating at a seeding rate of 120 trays per hour, the system achieved the accuracy of 99.19% for the single-seed seeding prediction, 94.79% for the reseeding prediction, and 93.43% for the missed seeding prediction, with an average computation time of 121 milliseconds per tray. This finding can also provide valuable support for the detection systems of the plug seedling seeding in leafy vegetables.