A satellite image target detection model based on an improved single-stage target detection network

Runwu Liu , Tian Wang (1) , Yi Zhou (2) , Chuanyun Wang (3) , Guangcun Shan (4) , Hichem Snoussi (5)

1 School of Automation and Electrical Engineering [Beijing] (University of Science and Technology Beijing)
2 Shanghai Jiao Tong University [Shanghai]
3 SAU - Shenyang Aerospace University
4 School of Instrumentation Science and Opto-electronics Engineering [Beijing]
5 LM2S - Laboratoire Modélisation et Sûreté des Systèmes

Aiming at the problem that it is difficult to detect small targets in satellite images, this paper proposes an improved method based on deep convolutional neural network YOLO V3. Firstly, the network structure of the original YOLO V3 was modified, and the target detection layer of three scales was reset. Then, during the detection process, since the test image is too large, the image is cut through the sliding window and then detected. During the experiment, the original YOLO V3 network and the improved network were used to train and test on the dataset. The experimental results show that the improved network improves the detection accuracy by 1.79% and the recall rate by 4.55%, the AP increased by 4.34%.

Domains: Automatic

Runwu Liu, Tian Wang, Yi Zhou, Chuanyun Wang, Guangcun Shan, et al.. A satellite image target detection model based on an improved single-stage target detection network. 2019 Chinese Automation Congress (CAC), Nov 2019, Hangzhou, China. pp.4931-4936, ⟨10.1109/CAC48633.2019.8997495⟩. ⟨hal-02486759⟩