基于修正CycleGAN的声呐图像库构建方法研究

Modified CycleGAN based sonar image library construction

声呐图像数据集获取困难,导致很多水下工作无法正常开展,如水下目标检测与跟踪、声呐图像的超分辨等,因此构建充足的声呐图像数据库成为很多水下研究工作的重要前提条件。受光学图像与合成孔径雷达(Synthetic Aperture Radar,SAR)图像转换研究工作的启发,提出了基于CycleGAN实现声呐图像库的构建,即利用光学图像合成声呐图像,实现光学到声呐的图像风格迁移。通过对CycleGAN网络损失函数的改进,提高了声呐图像的合成效果。通过与Pix2Pix等图像风格迁移网络进行比较的实验结果证明,修正后的CycleGAN网络具有更好的图像风格迁移效果。最后用合成的声呐图像训练Mask RCNN目标检测网络,并用真实的声呐图像进行测试,训练得到的模型能够成功地检测出真实声呐图像中对应的目标,进一步验证了利用光学图像构建声呐图像库的有效性。

Due to the difficulty of obtaining sonar image datasets,many underwater works,such as underwater object detection and tracking,super-resolution of sonar images,cannot be carried out normally.So,the construction of sufficient sonar image library becomes an important prerequisite for many underwater research works.Inspired by the conversion research work of optical image and synthetic aperture radar(SAR)image,the construction of sonar image library based on CycleGAN model is proposed to synthesize sonar image with optical image and realize image style transfer from optical to sonar.By improving the loss function of CycleGAN network,the effect of sonar image synthesis is improved.The experimental results comparing with Pix2Pix image style transfer network show that the modified CycleGAN network has better image style transfer effect.Finally,the Mask RCNN detection network is trained with synthesized sonar images and tested with real sonar images.The model obtained from training can successfully detect the corresponding targets in the real sonar images,which further verifies the effectiveness of the construction method of sonar image library using optical images.