Abstract: In urban planning and garden landscape, the artificial lawn plays a role in beautifying the environment, but the breeding of all kinds of lawn weeds seriously damages the ornamental ability of the landscape lawn. The artificial identification of weeds is timeconsuming and laborious, which affects the subsequent weeding efficiency. Therefore, based on the research results of deep learning, this study improves the existing Retina-Net target detection model by extracting the target image feature information of the training set, adding multiscale receptive field, improving the soft pool layer and other methods to improve the ability of weed detection and species discrimination of the model, which is conducive to the efficient development of subsequent weeding work. The recognition rates of six kinds of weeds in the final experiment were 85.3%, 84%, 89.6%, 86.7%, 95.1% and 91.5% respectively. Compared with other mainstream target detection algorithms, the recognition rate is improved by 2.2% to 9.3% respectively.

Key words: Retina-Net, image processing, CNN, object detection, receptive field, soft pooling

摘要： 在城市规划与园林景观中，人工养护的草坪起到美化环境的作用，但是各类草坪杂草的滋生，严重损害景观草坪的观赏性。而人工分辨杂草费时费力，影响后续的除草效率。因此，借助深度学习的研究成果，对现有的Retina-Net目标检测模型进行针对性改进，通过提取训练集目标图像特征信息、增设多尺度感受野、改进软池化层等方式，提升模型的杂草检测和种类分辨的能力，有助于后续除草工作的高效展开。最终试验对6类杂草的识别率分别为85.3%，84%，896%，86.7%，95.1%，91.5%。相比较于其他主流目标检测算法，识别率分别提高2.2%~9.3%。

关键词: Retina-Net, 图像处理, 卷积神经网络, 目标检测, 感受野, 软池化