Abstract: Aiming at the problems of complex background of paddy field, small size of weed image at the seedling stage, inaccurate identification and positioning under field natural environment photography, an improved YOLOv8 convolutional neural network method for weed detection at the seedling stage was proposed. A dedicated dataset based on the PASCAL VOC dataset format was constructed. First, DenseNet in the network convolution process was added to better extract weed features and optimize the vanishing gradient problem. Then, CBAM（Convolutional block attention module）attention mechanism was added to improve the model’s sensitivity to small sizes. Finally, the WIOU（Weighted intersection over union） loss function was used to optimize the loss function in the original network and improve the positioning ability of the model to the detection target. In the experiment, the performance of the improved algorithm was compared with algorithms such as Faster R-CNN, SSD （Single shot multiBox detector） and the original YOLOv8. The results showed that the improved algorithm was significantly superior to other algorithms, achieving an average precision of 97% and a detection speed of 100.3 frames/s on the test set, respectively. This high-precision and rapid detection capability met the demand for rapid and accurate detection in precision agriculture. This algorithm provided important theoretical and technical support for mechanical equipment to quickly identify weeds during the seedling stage and accurately spray pesticides.

Key words: YOLOv8, convolutional neural network, seedling weeds, target detection