High Imaging Quality Underwater Ghost Imaging Reconstruction Algorithm Based On Generative Adversarial Network

Correlation imaging,also known as"ghost imaging",is widely used in the field of underwater active optical imaging due to its high imaging resolution,strong anti-noise performance,long imaging distance,and non-locality.However,the imaging quality of ghost imaging under low sampling rate conditions will be greatly reduced.The contradiction between imaging efficiency and imaging quality has become the first key problem restricting the development of underwater ghost imaging.At the same time,compared with ghost imaging in free space,water will have a serious negative impact on the imaging quality of ghost imaging.Therefore,how to reduce the impact of water on the imaging quality of ghost imaging has become the key to the development of underwater ghost imaging.This article mainly focuses on the research of underwater ghost imaging technology,and proposes solutions to the problem of mutual constraint between imaging efficiency and imaging quality under complex underwater conditions.The main work of this article is summarized as follows:（1）In order to reduce the impact of complex underwater environments on the quality of underwater ghost imaging based on deep learning proposed in this paper,a large amount of"clear"and"fuzzy"underwater paired datasets is required for training.Although there are some open-source underwater image datasets,there is a lack of corresponding clear datasets,which makes it impossible to perform supervised training of the underwater ghost imaging reconstruction network model based on deep learning.To solve this problem,this paper proposes an underwater paired dataset generation algorithm based on Cycle-GAN.After training,two generator network models can be obtained:G_B2A^（C）and G_A2B^（C）.G_B2A^（C）can convert underwater fuzzy images into clear underwater datasets,and G_A2B^（C）can convert underwater clear images into fuzzy underwater datasets.Finally,in the simulation analysis,the generated dataset is classified and tested using the VGG16 classification network,quantitatively proving that the generated dataset is close to the real underwater image and meets the training requirements.（2）Complex environmental factors such as underwater suspended particles can affect the quality of underwater ghost imaging,resulting in poor imaging quality,especially under low sampling rate conditions.Therefore,how to perform high-quality reconstruction of underwater targets under complex underwater environments with low sampling conditions becomes a key issue to be addressed.To solve this problem,this paper proposes an underwater ghost imaging algorithm based on generative adversarial networks.The method uses a generative adversarial network model.The generator uses U-Net as the main network architecture and adds double skip connections and attention gate modules between corresponding layers to improve the reconstruction quality.The network is trained using the underwater paired datasets created in Chapter 2 to reduce the impact of complex underwater environments on the reconstruction quality.During network training,adversarial loss,perceptual loss,and pixel loss are incorporated into the total loss function with different weights to improve the quality of the reconstructed image.（3）Although using Cycle-GAN technology can successfully obtain"clear"and"fuzzy"underwater paired datasets,there are two problems.（1）The training of Cycle-GAN requires a large number of underwater fuzzy datasets,which cannot be directly obtained in real life.（2）The fuzzy degree of the fuzzy classes generated by Cycle-GAN is unitary.To solve these problems,an underwater paired dataset generation algorithm based on FUNIT is proposed.After training,two generator network models can be obtained:G_B2A^（F）and G_A2B^（F）.G_B2A^（F）can generate clear underwater datasets,and G_A2B^（F）can generate datasets with different blurriness degree.To further improve the reconstruction quality of underwater ghost imaging under low sampling conditions for underwater targets under complex underwater environments,an underwater ghost imaging algorithm based on few-shot learning is proposed.This method consists of two parts:the reconstruction part（UGI-DL）and the deblurring part(G_B2A^（F）).UGI-DL uses a generative adversarial network model for underwater target reconstruction.The generator of UGI-DL uses U-Net as the main network architecture and adds double skip connections and attention gate modules between corresponding layers to improve the reconstruction quality.Residual networks are also added between adjacent layers.The deblurring part further improves the reconstruction quality of the reconstructed results through G_B2A^（F）based on the reconstruction part.