Gaussian Wasserstein distance based ship target detection algorithm

Abstract

In remote sensing images, ship images are characterized by small target size, irregular arrangement and large scale difference, and the detection accuracy of existing target detection algorithms for small targets of ships is low. In order to extract the information in remote sensing target images more effectively and improve the detection accuracy of small targets, four models are constructed on the framework of GWD based on the dataset of HRSC16 in this paper. Firstly, the initial model based on GWD is trained, secondly, a cutout data enhancement strategy is added to GWD to further improve the robustness and overall performance of the convolutional neural network in the training process, then a ConvNeXt backbone network is introduced on the basis of GWD to improve the scalability and convergence of the network, and finally, the third model is based on the network is deepened, which can extract small target features more fully. The image recognition accuracy in HRSC16 dataset is improved from 54.1% at the beginning to 61.65%. Experiments show that the target detection algorithm proposed in this paper has good detection accuracy for small target detection of ships, and it is improved by 7.55% for small target detection after the improvement.