A Object detection Method for Missile-borne Images Based on Improved YOLOv3

Detecting small objects in complex circumstances is an important topic in the research of today' s object detection [1], especially in military, which needs more reliable, stable and accurate detection results. In order to improve the detection of small objects, we improved the structure of the YOLOv3 network by replacing the convolution module in the original network with multi-branch scale convolution, increasing the adaptability of the network to different sizes of objectss and reducing the number of network layers to balance the depth and width of the network, while also improving the feature extraction and representation capabilities. And based on the premise of a small number of data sets, we simulate some complex environments, which are composed of different weather, illumination, motion and rotational blur. We also enhance and extend the data in the network learning. Through the system simulation experiment, small objects can be recognized in such complex environments, which provides a reference for object detection of missile-borne images.