Abstract
Accurate assessment of the planting effect is crucial during the potato cultivation process. Currently, manual statistical methods are inefficient and challenging to evaluate in real-time. To address this issue, this study proposes a detection algorithm for the potato planting machine's seed potato scooping scene, based on an improved lightweight YOLO v5n model. Initially, the C3-Faster module is introduced, which reduces the number of parameters and computational load while maintaining detection accuracy. Subsequently, re-parameterized convolution (RepConv) is incorporated into the feature extraction network architecture, enhancing the model's inference speed by leveraging the correlation between features. Finally, to further improve the efficiency of the model for mobile applications, layer-adaptive magnitude-based pruning (LAMP) technology is employed to eliminate redundant channels with minimal impact on performance. The experimental results indicate that: 1) The improved YOLOv5n model exhibits a 56.8% reduction in parameters, a 56.1% decrease in giga floating point operations per second (GFLOPs), a 51.4% reduction in model size, and a 37.0% reduction in Embedded Device Inference Time compared to the YOLOv5n model. Additionally, the mean average precision (mAP) at mAP@0.5 achieves up to 98.0%. 2) Compared with the YOLO series model, mAP@0.5 is close, and the parameters, GFLOPs, and model size are significantly decreased. 3) Combining the ByteTrack algorithm and counting method, the accuracy of counting reaches 96.6%. Based on these improvements, we designed a potato precision planter metering system that supports real-time monitoring of omission, replanting, and qualified casting during the planting process. This system provides effective support for potato precision planting and offers a visual representation of the planting outcomes, demonstrating its practical value for the industry.