摘要
农作物的精确分割是实现智慧农业的关键技术之一,针对复杂环境下农业图像分割准确率低的问题,提出了一种基于边缘引导的特征融合网络(Boundary Guided Feature Fusion Network,BGFFNet)。首先,采用色彩抖动对农业图像进行预处理;然后,构建基于多尺度的边缘感知(Edge-Aware Module,EAM)编码器来提取农业图像中边缘、轮廓等多尺度细节信息,并将该信息作为解码器的先验表示,同时,构建边缘特征引导的聚合模块(Edge-guidance Feature Module,EFM)将来自编码器中的边缘先验与深层特征有效聚合;最后,采用交叉熵和dicceloss混合损失函数来缓解数据类别不平衡问题。实验结果表明:与传统的UNet分割模型相比,所提方法的DSC提高了6%、IoU提高了7.04%、TPR提高了9.7%。
Accurate segmentation of crops is one of the key technologies for the realization of smart agriculture.Addressing the issue of low accuracy in agricultural image segmentation under complex environments.Proposed an Boundary Guided Feature Fusion Network(BGFFNet).Firstly,agricultural images are pre-processed using color jittering.Then,an Edge-Aware Module(EAM)encoder based on multi-scale is constructed to extract multi-scale detail information such as edges and contours in agricultural images,and this information is used as the prior representation for the decoder.At the same time,an Edge-guidance Feature Module(EFM)is constructed to effectively aggregate the prior representation from the encoder with deep features.Finally,a hybrid loss function combining cross-entropy and dice loss is used to mitigate the issue of class imbalance in the data.The experimental results show that compared with the traditional UNet segmentation model,the DSC of the method proposed in this paper is increased by 6%,IOU is increased by 7.04%,and TPR is increased by 9.7%.
作者
吴德科
徐坤财
惠芸
王孟孟
WU Deke;XU Kuncai;HUI Yun;WANG Mengmeng(Public Education Department,Guiyang Institute of Information Science and Technology,Guizhou 550025,China;School of Intelligent Engineering,Guiyang Institute of Information Science and Technology,Guizhou 550025,China)
出处
《机械工程与自动化》
2025年第4期33-37,共5页
Mechanical Engineering & Automation
基金
贵州省青年科技人才成长项目(黔教技[2024]277,黔教技[2024]279)。
关键词
边缘引导
特征融合
色彩抖动
高斯滤波
边缘感知
edge-guidance
feature fusion
color jittering
gaussian filtering
edge awareness
