.png){kind=icon}
.png){kind=icon}
Return
Sonar images generation method based on improved stable diffusion combined with ControlNet model
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-026-02334-6.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With development of sonar technology and deep learning, sonar images have been widely used to detect underwater targets. However, the shortage of sonar data sources has become one of the biggest challenges in underwater targets detecting with neural networks. For example, actual underwater experiments are complicated and high cost, and the un-known detection targets and wreck degree, which makes it challenging on sonar object detection and recognition algorithms. In order to solve these problems, we propose a sonar image generating method based on diffusion model combined with ControlNet and LoRA method. The sonar-style target sample images are generated through Stable Diffusion with real sonar images. By combining the optical targets and sonar targets, the proposed image can produce multiple sonar-style images. Use of LoRA method with few real sonar images, which can train sonar stylized model to generate images. In the meantime, by introducing fine-tuned ControlNet model via control wreck degree of optical image, which makes the specific target be expanded with more diversification. The model makes the degree of damage to sonar target wreckage controllable. The experimental results show that our method achieves features similar to those of real sonar targets and can be used for training detection tasks.
Keywords:
Sonar images
Sonar-style target
Stable diffusion
ControlNet
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
No organization information available