An improved small target detection method based on Yolo V3

Abstract

Aiming at the efficiency and accuracy of small target detection in current traffic flow, this paper proposes an improved Yolo V3 method and applies it to small target detection. The method is to first optimize the network structure of Yolo V3, and add a new small target-friendly 4-fold down-sampling residual between the second residual block and the third residual block of Darknet-53 Block, improve the detection accuracy of small targets; perform 2-fold up sampling on the 8-fold down-sampling feature map output by the original network, and perform the 2-fold up-sampling feature map with the feature map output by the newly added third residual block Splicing, build a feature fusion target detection layer whose output is 4 times down sampling. The improved Yolo V3 algorithm is compared with the unimproved algorithm, and it is concluded that the improved algorithm can significantly improve the recall rate of small target detection and the average detection accuracy.