Abstract: In the complex background of sea clutter and various interfering objects, the traditional radar image object detection method has a low detection precision and high false alarm rates. The proposed deep learning method of the current hotspot research, is introduced into radar image target detection. Firstly, the advantages of the current advanced YOLOv3 detection algorithm and the limitations applied to the radar image field are analyzed, and the ship target detection data set with interference objects in the sea clutter environment is constructed. The data set contains various backgrounds, resolution, target position relations and other conditions, which can meet the actual task needs more completely. In allusion to the target sparseness and small target size, the Kmeans algorithm is used to calculate the anchor coordinates suitable for the dataset. Secondly, based on YOLOv3, an improved multiscale feature fusion prediction algorithm is proposed, which fuses the multilayer features information and the spatial pyramid pooling. Through a large number of comparative experiments, the mAP of the proposed method is improved by 6.07% compared with the original YOLOv3.

Key words: deep learning, radar image, object detection, YOLOv3