| Cite this article: | 刘尚旺,张杨杨,蔡同波,等.基于改进PSPnet的无人机农田场景语义分割[J].灌溉排水学报,0,():-. |
| LIU Shangwang,ZHANG Yangyang,CAI Tongbo,et al.基于改进PSPnet的无人机农田场景语义分割[J].灌溉排水学报,0,():-. |
|
| |
|
|
| This article has been：Browse 1133Times Download 0Times |
|
|
|
| DOI： |
|
| Semantic segmentation of UAV farmland scene based on improved PSPnet |
|
LIU Shangwang, ZHANG Yangyang, CAI Tongbo, TANG Xiufang, WANG Changgeng
|
|
Henan Normal University
|
| Abstract: |
| 【Background】Deep Learning, an important method in Artificial Intelligence, has further promoted the development of research in many fields. But the application in agriculture is still less and simple. On the other hand, UAV(Unmanned Aerial Vehicle) technology , which character for more flexible, more efficiency and low cost, has become a main method in the collecting of remote sensing information. It plays an important role in the crop classification, agricultural yield estimation, food security detection, agricultural water planning and so on. In order to get more effective interpretation of farmland images, this paper carried out the research of semantic segmentation of UAV farmland scene by the condition of deep learning and UAV technology.【Objective】As a semantic segmentation model, PSPnet has excellent segmentation performance, but the segmentation effect of UAV farmland scene still needs to be improved. To make PSPnet more suitable for UAV farmland scene semantic segmentation, there are three aspects employed to improve it in this paper. 【Method】Firstly, different dimensional features were concatenated to preserve more detailed image features via enhanced scene analysis. Secondly, a lightweight semantic segmentation model was built by using the deep separable convolution module, which is more efficient. Finally, the activation function was replaced to improve the segmentation effect of our model.【Result】 Experimental results show that the mean pixel accuracy and mean intersection over union of our method are 89.48% and 82.36%, respectively; compared with the original model, the segmentation results are accordingly improved by 18.12% and 18.93%; furthermore, our segmentation results are better than Unet and Deeplabv3+.【Conclusion】The above can prove that our model can effectively segment the farmland scene remotely sensed by UAV. |
| Key words: PSPnet, semantic segmentation, Feature concatenate, Deep separable convolution, Activation funciton |
|
|
|
|
|
|
