Research And Design Of Underwater Optical Image Restoration Method Based On Generative Adversarial Network

Underwater optical images are widely used in the exploration of underwater resources.The acquisition equipment of underwater optical images and the susceptibility to the occlusion of underwater organisms and underwater transmission noise cause the missing content of the finally obtained optical images,which adversely affects the subsequent analysis of underwater optical images.Therefore,in order to obtain complete underwater optical images,it is of great significance to research and design related image restoration algorithms.With the rapid development of deep learning technology,a series of optical image repair algorithms based on generative adversarial networks have been widely used in underwater optical image repair and have achieved certain repair effects.However,there are still problems that need to be improved in this series of methods: first,when the missing pixels in the underwater optical image have a low similarity with the surrounding pixels immediately adjacent to the missing area,the algorithm's repair effect is not good;second,only using The single-level network model is easy to make the semantic continuity of some areas of the repaired optical image low.Third,in the process of optimizing the underwater image repair model,it is difficult to improve the accuracy of deep learning algorithms.Easy sample balance method affects the ability of deep learning models to repair damaged pixels.Aiming at the above problems,this paper makes in-depth research on optical image repair theory and related repair methods,and performs the following work:1.Aiming at the problem of low similarity between missing pixels and neighboring intact pixels,this paper proposes an adaptive convolution algorithm based on generative adversarial networks.This algorithm introduces an attention strategy,which can adaptively analyze the correlation between missing pixels and intact pixels in the optical image and perform position shift to alleviate the constraint of local correlation a priori on generating adversarial networks.To verify the effectiveness of the algorithm,relevant comparative experiments were performed.2.Aiming at the problem of semantic continuity of the optical image after repair using only a single-level network,this paper proposes a method of network fusion.This method uses a cascading model strategy.After a single-stage generation adversarial network model is obtained,the generation model is re-constructed to reduce the area of the region with lower pixel continuity.To verify the effectiveness of the algorithm,relevant comparative experiments were performed.3.Aiming at the problem of lack of related easy and easy sample balance methods,this paper proposes a difficult case mining method suitable for generating adversarial networks.The method continuously calculates the difference between the current optical image repair result and the intact optical image during the training process,and adaptively performs weight balancing based on the magnitude of the difference.To verify the effectiveness of the algorithm,relevant comparative experiments were performed.