UAVs rely on scene understanding and autonomous movement to perform tasks. The common method of scene understanding is to use GPS + map. We hope to create offline maps through semantic segmentation to get rid of the constraints of GPS. But the UAV image scale changes too much. Related algorithms, for example, Multi-Scale-Dilation Net extracts multi-resolution features for fusion, and U-Net fuses features of various scales in the convolution process. The larger the scale of these two algorithms, the greater the weight of the target. This paper hopes to strengthen the characteristics of small-scale targets and proposes the algorithm Deep-Skip Net, which itself is also an encoder-decoder network similar to U-Net, but does not perform full-scale Skip connections. And Skip connection to the smallest two scales to ensure network accuracy. For the original scale, deep-level skip connections are used to retain the effective characteristics of small-scale objects. Deep-Skip Net conducted a comparative experiment on the UAV data set UAVid, and evaluated the results by averaging the intersection and comparison. The results of the comparative experiment show that, compared with the Multi-Scale-Dilation Net of the author of the data set, the overall performance of Deep-Skip Net is increased by 8%, and the segmentation effect for small-scale humans is increased by 20.4%. Both large-scale objects and small-scale objects have more balanced prediction capabilities.
Deep-Skip Net: Research on UAVs Image Semantic Segmentation Technology
Lect. Notes Electrical Eng.
International Conference on Autonomous Unmanned Systems ; 2021 ; Changsha, China September 24, 2021 - September 26, 2021
Proceedings of 2021 International Conference on Autonomous Unmanned Systems (ICAUS 2021) ; Kapitel : 23 ; 217-224
2022-03-18
8 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
British Library Conference Proceedings | 2005
|