English

中国农机化学报

中国农机化学报 ›› 2023, Vol. 44 ›› Issue (9): 154-160.DOI: 10.13733/j.jcam.issn.2095-5553.2023.09.022

• 农业信息化工程 • 上一篇    下一篇

基于生成对抗网络的田间杂草图像超分辨率重建

何朝霞1,朱嵘涛1,徐俊英2   

  1. 1. 长江大学文理学院,湖北荆州,434023; 2. 长江大学农学院,湖北荆州,434023
  • 出版日期:2023-09-15 发布日期:2023-10-07
  • 基金资助:
    国家青年科学基金项目(62101114);湖北省教育厅科学研究计划指导性项目(B2022474)

Superresolution reconstruction of images of weeds in the field based on generative adversarial network

He Zhaoxia1, Zhu Rongtao1, Xu Junying2   

  • Online:2023-09-15 Published:2023-10-07

摘要: 针对目前图像超分辨率方法生成精度以及性能方面的不足,以SRGAN(SuperResolution Using a Generative Adversarial Network)模型为基础,探讨和构建一种生成对抗网络模型,在该模型的损失函数中引入卷积神经网络边缘检测损失,生成的高分辨率图像中保持图像更多的细节。以作物/杂草田地图像数据集作为测试集,开展一系列试验,本方法超分辨率生成图像较双三次插值、SRGAN、ESRGAN和深度残差的GAN模型,峰值信噪比PSNR分别高8.242dB、5.521dB、3.079dB、2.339dB;结构相似性SSIM分别高0.143、0.089、0.051、0.018;在AI识图中的识别准确率分别高10.6%、3.5%、3.9%、2.7%。该方法为田间杂草图像的相关研究提供思路,同时为田间杂草分类等应用研究做好前期数据准备。

关键词: 超分辨率, SRGAN, 生成对抗网络, 损失函数, 边缘检测损失, 田间管理

Abstract: Superresolution reconstruction plays an important role in the field of intelligent agriculture. In view of the shortcomings in the generation accuracy and performance of the image superresolution recently, a generation adversarial network based on the model of SRGAN was discussed and constructed. In this model, the convolution network edge detection loss was introduced into the loss function, and more details of the image were maintained in the generated highresolution image. A series of experiments were carried out using sets of crop/weed field image data as test sets, compared with bicubic interpolation, SRGAN, ESRGAN and GAN model with depth residuals, the PSNR was 8.242dB, 5.521dB, 3.079dB and 2.339dB higher, the SSIM was 0.143, 0.089, 0.051 and 0.018 higher, and the recognition accuracy in AI was 10.6% , 3.5% , 3.9% and 2.7% higher, respectively. It provided ideas and methods for the related research of other images of weeds in the field, and prepared the preliminary data for application research of image classification of field weeds.

Key words: superresolution, SRGAN, generative adversarial network, loss function, loss of edge detection, field management

中图分类号: