关键词: 稻田杂草, 野慈姑, 语义分割, DeepLabV3+

Abstract: Paddy weed position acquisition is the basis of targeting spraying herbicide and mechanical intelligence weeding. In order to acquire information acquisition of weeds in seedling stage under natural light environment and complex paddy field background, in this paper, a new semantic segmentation method for weeds at seedling stage was proposed based on the full convolutional neural network and DeepLabV3+ was used for semantic segmentation of seedlings and weeds to obtain the weed location information by sagittaria trifolia as the research object. Firstly, the weeds RGB image of sagittaria trifolia in the paddy field were captured, and each pixel of the seedlings, weeds and background in the images were manually labeled by the image labeling tool LabelMe. 70% data set was used for the parameter training of the DeepLabV3+ network model, and 30% data set was used to test the performance of DeepLabV3+. By comparing with FCN and U-Net semantic segmentation methods, the proposed DeepLabV3+ semantic segmentation method had the best performance indicators, such as accuracy, proportional ratio, frequency weight intersection ratio and F value, and the experiment results showed that the pixel accuracy of DeepLabV3+ model was up to 922%. The accuracy rates of U-Net and FCN methods were 92.1% and 84.7% respectively. The method proposed in this paper could accurately segment weeds, seedlings and background pixels at seedling stage of paddy field, and also meet the practical application requirements of intelligent weeding and targeted herbicide spraying.

Key words: paddy weed, sagittaria trifolia, semantic segmentation, DeepLabV3+