基于Transformer结构的遥感影像敏感目标自动隐藏方法

遥感影像敏感目标隐藏是保证遥感资源安全共享的关键。针对传统方法存在的目标检测不完全、补全结果不可靠的问题，提出了一种基于Transformer结构的遥感影像敏感目标自动隐藏方法。首先利用以Swin Transformer为主干网络的Cascade Mask R-CNN（region-based convolutional neural network）实例分割优化模型检测敏感目标并生成掩膜区域，同时设计了RSMosaic（remote sense Mosaic）合成数据方法减少人工标注数据；然后，基于色相-饱和度-明度（hue-saturation-value, HSV）空间的阴影检测模型扩展掩膜区域；最后，引入MAE（masked autoencoders）模型实现目标背景生成。以飞机目标为例，与Partial-Connvolutios和EdgeConnec进行了对比实验。结果表明，相比传统方法，该方法在敏感目标实例分割中的边界框与像素掩膜AP值分别提升了13.2%与11.2%；在使用RSMosaic合成数据后，边界框与像素掩膜AP值可分别再提升9.39%与14.16%, 且图像修补中的平均绝对误差和最大平均差异提升80%以上，实现了结构合理、纹理清晰的敏感目标自动隐藏效果。

Automatic Hiding Method of Sensitive Targets in Remote Sensing Images Based on Transformer Structure

  Objective  Decryption is the key to ensure the safe sharing of remote sensing resources. To solve the problems of incomplete target detection, unreliable complementary results, high resource consumption and difficulty of training in the traditional methods of sensitive target hiding in remote sensing images, an automatic hiding method of sensitive targets in remote sensing images is proposed based on the ability of Transformer structure to deal with global information.  Methods  Firstly, the optimized Cascade Mask R-CNN instance segmentation model with Swin Transformer as the backbone network is used to detect sensitive targets and generate mask regions. After improving the generalization capability of the model, RSMosaic (remote sense Mosaic), a data synthesis method to reduce the dependence on manually labeled data is designed. Secondly, the mask region is expanded by using the shadow detection model based on HSV(hue-saturation-value) space, and the MAE(masked autoencoders) model is introduced to achieve target background generation. Finally, the generated images are spliced with the original images to obtain the decrypted images.  Results  The sub-meter remote sensing images collected by Google Earth are used as test data, and the results show that this proposed method generates reliable hiding results while reducing dataset dependence and training resource consumption. Compared with the traditional method, the AP (average precision) values of bounding box and pixel mask are improved by 13.2% and 11.2% respectively in sensitive target instance segmentation, and the AP values can be improved by another 9.39% and 14.16% respectively after using RSMosaic, which is better than other repair models in terms of objective index and index variance in the field of image repair, especially in mean absolute error and maximum mean discrepancy indexes which are improved by more than 80%. It achieves the effect of automatic hiding of sensitive targets with reasonable structure and clear texture.  Conclusions  The proposed method reduces manpower, data and computing resources, and achieves better results in both subjective visual effects and objective indexes, which can provide technical support for real remote sensing image sharing.