Maize Seedling and Weed Detection Using BFSL-YOLOv8

Abstract

Addressing the urgent need for weed control in maize seedlings due to the low detection accuracy in complex agricultural environments, this paper proposes a BFSL-YOLOv8 model. paper proposes a BFSL-YOLOv8 model. The model enhances multi-scale feature extraction and fusion by integrating a Spatial Pyramid Pooling with Large Separable Kernel Attention and skip connections (SPPF_LSKA) module into the neck of YOLOv8 and efficiently captures long-range dependencies by introducing an improved BiFormer module in the head. Experiments were conducted on a publicly available real-world field monitoring dataset under Experiments were conducted on a publicly available real-world field monitoring dataset under specific hardware and software environments, using stochastic gradient descent (SGD) for training. achieves a mean Average Precision (mAP) of 99.5% for maize detection and improves the mAP at an Intersection over Union (IoU) threshold of 0.5 (mAP50) for weed detection by 0.5 percentage points to 64.8% compared to the baseline YOLOv8. The model has 3.70M parameters and a processing time of 4.4ms per image. F1-confidence curve analysis indicates that BFSL-YOLOv8 exhibits good robustness in complex scenarios, with an optimal confidence threshold of 0.292 and an average confidence threshold of 1.5 times. of 0.292 and an average F1-score of 0.82 for all classes. The experiments validate the effectiveness of the proposed method, providing a new solution for accurate and efficient weed detection in maize. accurate and efficient weed detection in maize fields for precision agriculture.