Real-time monitoring method for peanut double-seed precision seeding quality based on machine vision and an improved YOLOv8

Double-seed hill planting is one of the important sowing modes for the effective seedling emergence in peanut cultivation. However, conventional monitoring techniques with photoelectric sensors cannot accurately identify the specific quantity of peanut seeds per hill, the extent of seed damage, or the inclusion of foreign debris. It is often required for high precision, real-time, and functionality in modern agricultural machinery. This study aims to investigate the real-time monitoring of peanut double-seed hill planting quality using machine vision and an improved YOLOv8 architecture. Specifically, a lightweight detection model was tailored for the real-time recognition of peanut quality, termed the MobileNetV3-SimAM-Focal_GIoU YOLOv8 (MSFG-YOLOv8). The MobileNetV3 network was integrated as the backbone to extract feature information from peanut seeds, indicating the optimal balance between detection accuracy and computational efficiency. Furthermore, the SimAM attention mechanism was introduced into the network architecture. Features were extracted from the stacked and damaged seeds, particularly for the occlusion and morphological irregularities. In addition, bounding box regression was used to further optimize the architecture. The loss function was adopted as the Focal_GIoU. Class balance was realized for the high localization precision on small and dense targets. A robust experimental platform was constructed for peanut double-seed hill planting. The Raspberry Pi 5 system served as the processing unit. A high-definition sampling camera was integrated with a human-machine interaction module. A configuration was embedded to validate the feasibility and performance of the algorithm in the field conditions. Extensive experiments were conducted to evaluate the performance of the MSFG-YOLOv8 model against the standard YOLOv8s baseline. The results demonstrate that the lightweight model achieved superior performance in significantly reducing computational resource consumption. Specifically, the MSFG-YOLOv8 model was improved Precision by 4.4 percentage points, recall by 5.0 percentage points, and the mean Average Precision (mAP) by 4.7 percentage points, compared with the original YOLOv8s. In the complexity and efficiency, the computational load (measured in FLOPs) was reduced by 42.8%, the parameter count decreased by 65.1%, and the overall model size was compressed to only 14.8 MB. Furthermore, the seeding quality indicators were determined to be highly suitable for the deployment of resource-constrained edge devices. MSFG-YOLOv8 model achieved Average Precision (AP) values of 99.6%, 94.4%, 99.8%, and 95.4%, respectively, for the qualified rate, over-seeding rate, missed-seeding rate, and damage rate. The high accuracy of the system attained a qualified rate of 99.8%, the multiple-seeding rate of 96.2%, and the missed-seeding rate of 96.9%. The improved MSFG-YOLOv8 model was constructed for the real-time tracking and statistical analysis of deep learning features from individual seeds in fields. A polygonal region counting module was integrated with the Multi-Object Tracker using Kalman filtering and the Hungarian algorithm. The workflow was as follows: Firstly, the MSFG-YOLOv8 model was used to detect the peanut seeding images for the positional and status information of each seed; Secondly, the information was then fed into the MOT. Kalman filtering was utilized to predict target motion states. The Hungarian algorithm was employed for the optimal matching between predicted and actual values. Interframe target mapping was realized after evaluation; Finally, the counting module triggered the counts when seed centroids were predefined within quadrilateral monitoring region boundaries, according to geometric spatial criteria. Performance monitoring metrics were obtained during seeding. The indicators were determined for the real-time perception of peanut sowing performance, such as qualified rate, over-seeding rate, and missed-seeding rate on the screen. The sowing status was then adjusted to significantly reduce the manual inspection. The system can provide reliable detection of seed count and integrity. The findings can provide a solid technical reference to maintain the high-quality standards in peanut double-seed hill planting.