Title: An improved object detection algorithm based on Depthwise Separable Convolutions 

Author(s): Yu, XY (Yu Xiuyuan); Bao, QL (Bao Qiliang); Jia, HL (Jia Haolong); Yu, L (Yu, Li); Rui, Q (Rui, Qin) 

Edited by: Tianran W; Tianyou C; Huitao F; Qifeng Y 

Source: SECOND TARGET RECOGNITION AND ARTIFICIAL INTELLIGENCE SUMMIT FORUM  Book Series: Proceedings of SPIE  Volume: 11427  Article Number: 114272T  DOI: 10.1117/12.2552710  Published: 2020   

Abstract: Aiming at small objects detection such as unmanned aerial vehicle (UAV), this paper proposes a fast object detection algorithm based on depthwise separable convolutions. Firstly, the inverted residuals units based on depthwise convolutions and pointwise convolutions are used to construct a lightweight feature extraction network to improve the network's speed. Secondly, the feature pyramid network is used to detect the five scale feature maps to improve the detection performance of small objects. Otherwise, we make an UAV dataset based on the urban background for training and testing of our experiments. The experimental results show that the improved method proposed in this paper can effectively improve the detection accuracy and real-time performance of UAVs in complex urban backgrounds, and the computation of network is greatly reduced, thereby making it possible to achieve object detection on embedded systems. 

ISSN: 0277-786X 

eISSN: 1996-756X 

ISBN: 978-1-5106-3632-3