Analysis of the Impact of Different Improvement Methods Based on YOLOV8 for Weed Detection

Abstract

In response to the issues of missed detection, false positives, and low recognition rates for specific weed species during weed detection, a YOLOv8-based improved weed detection algorithm named EDS-YOLOv8 is proposed. Improvements were made in three main aspects. First, the YOLOv8 backbone network was enhanced with EfficientViT and RepViT architectures to improve the detection capability of dense-type weeds. Second, different attention mechanisms were added, such as SimAM and EMA, to learn 3D weights and achieve full fusion of features. BiFormer was introduced for dynamic sparse attention and resource allocation. Third, significant module improvement involved introducing dynamic snake convolution into the C2f module to further enhance detection capabilities for deformable objects, especially needle-shaped weeds. The improved model is validated on the established weed dataset. The results show that combining the original backbone network with dynamic snake convolutions yields the highest performance improvement. Precision, recall, mAP (0.5), and mAP (0.5:0.95) are improved by 5.6%, 5.8%, 6.4%, and 1%, respectively, and ablation experiments on the effects of the three improvement methods on model performance show that using EfficientViT as the backbone network while simultaneously improving the crucial module and adding the SimAM attention mechanism effectively enhances the model’s performance. Precision, recall, mAP (0.5), and mAP (0.5:0.95) are improved by 6%, 5.9%, 6.4%, and 0.7%, respectively.