| Sonar image is the important basis for underwater detection research.However,the harsh underwater imaging environment and high sonar detection costs severely limit the quality of sonar images and the size of the dataset.Especially in environments where deep learning and other related methods are widely used,the size of sonar image datasets directly affects the learning effectiveness of deep learning and its application in underwater target detection and other fields.In addition,the scarcity of sonar image datasets has led to incomplete evaluation systems and methods,which hinders research on underwater related tasks to a certain extent.In response to the above issues,this article conducts in-depth research on the generation methods of sonar image datasets and the robustness evaluation methods of sonar image datasets.The main research contents are as follows:(1)A D3GAN(Deep Determined Double GAN)model is proposed to address the scarcity and insufficient scale of sonar image datasets,which combines deep reinforcement learning methods with generative adversarial networks.This model is mainly composed of two sets of generative adversarial networks combined with Reinforcement Learning method DDPG,which achieve style transfer from optical images to sonar images.The model utilizes the trial-anderror mechanism of Deep Reinforcement Learning method DDPG to supervise the direction of generating adversarial network generated images and guide the completion of supplementing important interpretation features such as shadow features.On this basis,the sonar image datasets generated by D3 GAN and Cycle GAN were compared and subjectively analyzed.D3 GAN can successfully and quickly converge,and the generated sonar image dataset is not inferior in image quality to Cycle GAN.Both models can achieve appropriate transfer of most features.Specifically,D3 GAN outperforms Cycle GAN in supplementing important interpretation features such as shadow features,and is more adaptable to the generation of sonar image datasets.(2)A multi-indicator normalized set evaluation method for sonar image datasets is proposed to address the issue of incomplete robustness evaluation system and methods.This method combines Generative Adversarial Network evaluation methods,image data quality evaluation methods,image robustness evaluation,and specific experimental task evaluation methods.It comprehensively evaluates the overall robustness of the sonar image dataset based on certain rules,including the image quality of the sonar image dataset,the difference between the generated sonar image and the real sonar image,the ability of the generated sonar image to resist robustness attacks,and its performance in specific recognition tasks.Based on the multiindicator normalization set evaluation method,the sonar image datasets generated by D3 GAN and Cycle GAN were evaluated separately.The comprehensive robustness of the sonar image datasets generated by D3 GAN model is about 24.17% higher than that of Cycle GAN.From an objective data perspective,it is shown that D3 GAN is more suitable for the generation of sonar image datasets compared to Cycle GAN. |
