Multi-scale ship target detection in remote sensing images

Abstract

In this paper, a new method is proposed for multi-scale ship target detection and recognition in remote sensing image, which uses the change receptive field design of neural network to realize multi-scale target detection. The network structure is composed of three branches, and the expansion convolution is used to realize the effective perception of different size targets. In the training stage, the three branches share the same value parameters, and the training samples are divided into three sizes: large, medium and small, and the three branches are trained respectively; during the testing stage, the target detection is completed on the three branches at the same time, and finally the detection results are synthesized (non-maximally suppressed, NMS) to obtain the final detection results. Through the experiments on the Google Earth remote sensing image data set collected by ourselves, it is proved that the method is effective in multi-scale ship target detection in remote sensing image, and the performance is improved compared with SSD, FPN and other classical algorithms.