基于双分支深度神经网络的农田场景语义分割方法

Dual branches deep neural network for semantic segmentation in agricultural scenes

农田场景下对农作物和杂草的精确定位是靶向喷施除草剂和机械智能除草等技术的基础,针对现有算法易受目标间相互遮挡、目标形变、环境光照变化等不利因素影响的问题,提出一种基于双分支神经网络的农田场景语义分割算法,实现对农作物和杂草的像素级分类进而获取农作物和杂草的位置信息。首先,设计基于ResNeSt网络结构的骨干网络对图像进行特征提取;然后,设计并行的由细节分支和语义上下文分支组成的双分支神经网络,分别用于提取图像细节信息和图像语义上下文类别信息,并引入注意力机制以更好的提取上下文特征,提升语义分割的性能;接着,使用双分支特征融合模块对上述细节分支和语义上下文分支输出的特征进行融合;最后,通过语义分割头模块输出对农作物和杂草的语义分割结果。在自建数据集上的试验表明,所提出的算法能够对农作物和杂草进行像素级的准确分割,mIoU值达到93.8%,能够满足智能除草和除草剂靶向喷施的实际应用需求。

Accurate localization of crops and weeds in agricultural field scenes is the foundation for targeted spraying of herbicides and mechanical intelligent weeding.To address the issues of mutual occlusion between targets and target deformation that existing algorithms are susceptible to,a semantic segmentation algorithm was proposed for agricultural field scenes based on a dual‑branch neural network.This algorithm achieves pixel‑level classification of crops and weeds,thereby obtaining their precise location information.Firstly,we designed a backbone network based on the ResNeSt architecture to extract features from input images.Then,we proposed a parallel dual‑branch neural network consisting of a detail branch and a semantic context branch.The detail branch focuses on extracting fine‑grained information from images,while the semantic context branch captures high‑level semantic contextual information.Attention mechanisms were introduced to better extract contextual features and enhance the performance of semantic segmentation.Next,we performed effective feature fusion using a dual‑branch feature fusion module to combine the features extracted from the detail branch and the semantic context branch.Finally,the semantic segmentation head module outputs the semantic segmentation results for crops and weeds.Experimental results on our self‑built dataset demonstrate that the proposed semantic segmentation algorithm for agricultural field scenes achieves pixel‑level accurate segmentation of crops and weeds,with an mIoU(mean Intersection over Union)value of 93.8%.This algorithm meets the practical application requirements of intelligent weeding and targeted herbicide spraying.