Small Aerial Target Detection Algorithm Based on Improved YOLOv5.

Small target detection is a difficult problem in the field of target detection. Small targets have many characteristics, such as less visual information, difficult to extract discriminative features, easy background confusion and so on, as a result, its detection accuracy is often only half of the large target. To solve these problems, we study small target detection in UAV aerial scene, and propose an improved YOLOv5-based small target detection algorithm. First, we performed two optimizations on the YOLOv5 prediction scale, reducing the missed detection of small targets and reducing model parameters and computational complexity, while reducing the negative impact of scale mismatches. Then, based on the idea of BiFPN, we introduce the skip connection from backbone network to PAN, and use Concat weighted fusion instead of the weighted fusion in BiFPN to improve the detection precision of the model. Finally, we do some experiments on the data set of Visdrone2019. The experimental results show that the [email protected] and [email protected] :0.95 of the improved model are improved by 6.6% and 4.2%, respectively, the parameters of the model are reduced and YOLOv5 is light weight.