Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (12): 217-227.DOI: 10.3778/j.issn.1002-8331.2203-0182

• Graphics and Image Processing • Previous Articles     Next Articles

Improved DeepLabv3+ Model UAV Image Farmland Information Extraction

CHEN Yuqing, WANG Xiuxin   

  1. 1.College of Computer Science and Engineering, Guangxi Normal University, Guilin, Guangxi 541004, China
    2.Guangxi Key Lab of Multi-Source Information Mining & Security, Guilin, Guangxi 541004, China
    3.Guangxi Collaborative Innovation Center of Multi-Source Information Integration and Intelligent Processing, Guilin, Guangxi 541004, China
  • Online:2023-06-15 Published:2023-06-15

改进DeepLabv3+模型无人机图像农田信息提取

陈雨情,王修信   

  1. 1.广西师范大学 计算机科学与工程学院,广西 桂林 541004
    2.广西多源信息挖掘与安全重点实验室,广西 桂林 541004
    3.广西区域多源信息集成与智能处理协同创新中心,广西 桂林 541004

Abstract: Since the implementation of the rural revitalization strategy, UAVs have played an important role in China’s smart agricultural production as a high-tech tool. However, the accuracy of information extraction is not hight, especially for the edge of farmland and small farmland. Aiming at this problem, an improved DeepLabv3+ model with GhostNet as the backbone network is proposed. This method strengthens feature extraction from the backbone network through feature pyramid network(FPN). Then, this paper replaces the convolution of the atrous spatial pyramid pooling module(ASPP) in the encoder with a spatially aware stand-alone self-attention layer and adjusts the dilated rate in the ASPP module to improve the extraction accuracy of the farmland edge. In order to further fuse multi-scale information, the decoder performs the same operation as described above. Finally, without reducing the performance of the model, the Concatenate method is replaced with Add to reduce the training parameters of the model. The experimental results show that the mean intersection over union (mIOU) and mean pixel accuracy (mPA) of improved DeepLabv3+ model are 94.57% and 97.16%, which are 4.53 and 2.93 percentage points higher than DeepLabv3+ model respectively, which effectively improves the accuracy of farmland information extraction of farmland edge and small farmland.

Key words: UAV, farmland information, extraction, improved DeepLabv3+, semantic segmentation

摘要: 自乡村振兴战略实施以来,无人机作为一种高科技工具为我国智慧农业生产发挥着重要作用。但存在信息提取的精准度不高的问题,特别是对农田边缘和小农田的信息提取精度不高。针对该问题,提出一种以GhostNet为骨干网络的改进DeepLabv3+模型的研究方法。该方法将从骨干网络中提取的特征通过特征金字塔网络(FPN)加强特征提取;将编码器中空洞空间金字塔池化模块(ASPP)的1×1卷积替换成空间感知独立自注意层并将ASPP模块中的扩张率进行一定调整,以提高农田边缘的提取精度;为进一步融合多尺度信息将解码器进行上述同样操作;在不降低模型性能的前提下,将特征堆叠(Concatenate)用特征融合(Add)进行替换,以减少模型的训练参数。实验结果表明,改进DeepLabv3+模型平均交并比(mIoU)可达94.57%,平均像素精度(mPA)可达97.16%,相比于DeepLabv3+模型分别提高了4.53%和2.93个百分点,有效提高了农田边缘和小农田的信息提取精度。

关键词: 无人机, 农田信息, 提取, 改进DeepLabv3+, 语义分割