A Study on Weed Detection Based on Improved Yolo v5

Currently, weeds in agricultural fields are one of the most common and serious biological threats to agricultural production. These weeds have a strong ability to reproduce, and they directly or indirectly affect crop yield and quality by competing with crops for nutrients, causing crop diseases, and attracting noxious insects, thus bringing great losses to agricultural production. In this aspect of weed detection Zhai et al. proposed a crop row recognition algorithm based on Census transform, and the experiment was only carried out on the cotton field video with simple background, and its correct rate was 92.58%. However, the algorithm requires a higher experimental environment because it did not go to the actual cotton field environment for detection.2021, Kaijing Li et al. proposed a YOLOv3 cotton field weed detection algorithm based on YOLOv3 for the complex cotton field environment in Xinjiang, and took the weeds near the cotton seedling during the moving process as the detection target, and trained the improved YOLOv3 network model through the improvement of the feature extraction network. Although there is some success, the accuracy of weed detection is still not high in complex situations. To solve this problem, this paper proposes a new model yolo v5-ct based on Yolo v5 by introducing the cbam module and the Transformer encoder module in the backbone feature extraction network, which improves the prominence of weeds in complex farmland backgrounds and strengthens the network's focus on weeds. The experimental results show that in relatively complex environments, the yolo v5-ct model effectively improves the problem of low weed recognition accuracy in complex environments.