Military Target Detection Method Based on Improved YOLOv3 Network

Yu Bowen, Zhang Jie

target detection, deep learning, YOLOv3, complex scenes, situational awareness, residual network, dense connection network.

Improving the performance and accuracy of image target detection technology is an effective means to improve the generation and analysis ability of battlefield situational awareness. A training data set is constructed for complex battlefield environment, which contains relevant data conforming to the conditions of small targets, occlusion and relatively dense, etc., and can provide a test environment for various target detection algorithms. A military target detection method based on convolutional neural network RDBN-YOLOv3 algorithm is proposed to improve the efficiency and accuracy of military target detection in complex environments. Based on the characteristics of residual network and dense connection network, the residual dense connection structure is proposed and RDBN-YOLOv3 network structure is designed. The dense residual connection structure improves the fusion and reuse capability of the original YOLOv3 network for feature information at all levels. By combining local residual learning, global residual learning and global feature fusion strategy, the transmission of image feature information is optimized, and the detection performance of small targets, occlusion and relatively dense military targets is improved. Finally, experiments are carried out in the data set constructed in this paper. The experimental results show that compared with the original YOLOv3 algorithm, the average accuracy is 4.82% higher, which can provide effective technical support for battlefield situation generation and analysis.

https://dx.doi.org/10.31873/IJEAS.8.9.09