Object Detection in Optical Remote Sensing Images Based on Improved Lightweight Neural Network

Abstract

The optical remote sensing images collected by Unmanned Aerial Vehicle Remote Sensing (UAVRS) with real-time information, and object detection of the optical remote sensing images has significant development potential in the many fields such as transportation and agriculture. In addition to large objects such as buildings, small objects such as vehicles and ships can also be clearly observed in the collected high-resolution remote sensing images. This paper mainly focuses on the detection of vehicles and ships in remote sensing images, and proposes Scene-SSD based on the main principles of MobileNetV3 and SSD. In this paper, we improve the basic block bottleneck of MobileNetV3, introduce Generalized Focal Loss (GFL) function to replace the original loss function in SSD, improve the class imbalance problem and make the bounding box estimations are more precise, and the network model is trained by transfer learning to improve its generalization ability. It is experimentally illustrated that in object detection of remote sensing images, the Scene-SSD proposed in this paper is fast and the tested mAP can reach 77.9%, which is better than the MobileNetV3-SSDLite with the same network structure in the comparison test.