Dual-channel coupled target detection algorithm with dense scene

In the field of target detection in air-to-ground scenarios, the traditional single-stage detection algorithm has a poor effect on large-resolution image detection due to fixed-scale input, especially when there are more dense and small targets in the image. The phenomenon of missed detection is serious. Therefore, imitating the target search behavior of the human eye, a two-channel coupled target detection algorithm focusing on dense scenes is proposed. Based on the You Only Look Once V3(YoloV3) network, the algorithm adds dense scene detection channels to detect dense areas in the image. The scene coupling structure is established to fuse the feature information of the dense scene channel with the information of the target instance detection channel. Difficult area variable resolution detection is used to improve the detection accuracy of dense and small targets. The algorithm is validated by self-made air-to-ground dense scene dataset. Experiments show that the proposed algorithm has better effect on the detection of dense and small targets. Compared with the traditional YoloV3 network, the average accuracy is increased by 16.4% when the detection speed drops by 9.1 frames/s. © 2020, Editorial Department of Journal of Chinese Inertial Technology. All right reserved.