可见光遥感图像船舶目标数据增强方法研究

The lack of ship target samples in visible remote sensing image leads to the imbalance between positive and negative sample classes in training process. The lightweight neural network sparse MobileNetV2 ship target detection model obtained in training is easy to fall into overfitting. In this study, based on the generative adversarial networks and the image operation in the sample space, a ship target data augmentation method is proposed to realize the supplement of ship target, which combines the feature space and sample space (SADA). According to the characteristics of target size and gray distribution in remote sensing image, a deep convolution ship target generative adversarial network is established. The methods of sample data transformation and feature space fitting are fused in dada set layer to augment the training set. The experiments were conducted adopting the 0.5m resolution image data set screened in Google Earth. Experiment results show that the sparse MobileNetV2 network trained by the augmentation data set can increase the recall rate of ship target detection by 68.5%, which proves the feasibility of this method in improving the accuracy of lightweight deep learning model in small sample size visible remote sensing image ship target detection application.