Abstract Weed control is a global issue, and has attracted great attention in recent years. Deploying autonomous robots for weed removal has great potential in terms of constructing environment-friendly agriculture, and saving manpower. In this paper, we propose a weed/crop segmentation network that provides better performance for precisely recognizing the weed with arbitrary shape in complex environment condition, and offers great support for autonomous robots to successfully reduce the density of weed. Our deep neural network (DNN)-based segmentation model obtains persistent improvements by integrating four additional components. i) Hybrid dilated convolution and DropBlock are introduced into the classification backbone network, where the hybrid dilated convolution enlarges the receptive field, while DropBlock regularizes the weight parameters to learn robust features by random drops contiguous regions. ii) A universal function approximation block is added to the front-end of the backbone network, which adaptively converts the existing RGB-NIR bands into optimized (RGB + NIR)-based indices to increase the classification performance. iii) The bridge attention block is exploited, in order to make the network “globally” refer to the correlated region, regardless of the distance for capturing the rich long-range contextual information. iv) The spatial pyramid refinement block is inserted to fuse multi-scale feature maps with different size of receptive fields to provide the precise localization of segmentation result, by maintaining the consistency of feature maps. We evaluate our network performance on two challenging Stuttgart and Bonn datasets. The state-of-the-art performance on the two datasets shows that each added component has notable potential to boost the segmentation accuracy.

中文翻译：

---

一种用于检测杂草和作物的基于 DNN 的语义分割

摘要 杂草控制是一个全球性问题，近年来受到了广泛关注。部署自主除草机器人，在建设环境友好型农业、节约人力等方面具有巨大潜力。在本文中，我们提出了一种杂草/作物分割网络，该网络为在复杂环境条件下精确识别任意形状的杂草提供了更好的性能，并为自主机器人成功降低杂草密度提供了极大的支持。我们基于深度神经网络 (DNN) 的分割模型通过集成四个附加组件获得持续改进。i) 在分类骨干网络中引入混合扩张卷积和 DropBlock，其中混合扩张卷积扩大了感受野，而 DropBlock 正则化权重参数以通过随机丢弃连续区域来学习鲁棒特征。ii) 在骨干网络的前端添加了一个通用函数逼近块，它自适应地将现有的 RGB-NIR 波段转换为基于 (RGB + NIR) 的优化指标，以提高分类性能。iii) 利用桥注意力块，使网络“全局”引用相关区域，而不管捕获丰富的远程上下文信息的距离。iv) 插入空间金字塔细化块以融合具有不同感受野大小的多尺度特征图，通过保持特征图的一致性来提供分割结果的精确定位。我们在两个具有挑战性的斯图加特和波恩数据集上评估我们的网络性能。两个数据集上的最新性能表明，每个添加的组件都具有提高分割精度的显着潜力。