Due to real-world constraints, the quantity of ship datasets derived from remote sensing data is substantially limited and can’t fulfill the extensive sample demands required for training deep learning algorithms. According to this problem, a high-quality synthesizing method for three-band optical high-resolution remote sensing images containing ship targets is introduced, which utilizes 3D models and generative adversarial networks with style transfer capabilities. Based on the constructed dataset, synthetic samples are generated and evaluated. The experiments indicate that the approach can synthesize images visually close to real images. Incorporating these synthetic samples into the training process of detection models results in an increase of 2.6% in mAP for Faster R-CNN and 2.3% for YOLOv5.

ship detection, high-resolution optical image, synthetic samples, deep learning, generative adversarial networks

基于遥感数据获取的真实舰船数据集数量非常有限, 难以满足深度学习算法训练对样本数量的需求。针对此问题, 利用三维模型和能够进行风格迁移的生成对抗网络, 提出一种高质量的包含舰船目标的三波段光学高分辨率遥感图像仿真方法。基于构建的数据集, 进行仿真样本的生成及评估。研究结果表明, 该方法能够合成在视觉上接近真实影像的图像, 通过加入合成样本对目标检测模型进行训练, 可以使Faster-RCNN和YOLOv5的全类平均正确率mAP分别提升2.6%和2.3%。

舰船目标检测, 高分辨率光学影像, 仿真样本, 深度学习, 生成对抗网络