Research On Underwater Image Enhancement Algorithm Based On Deep Learnin

As the main carrier of underwater information,underwater images play an important role in human exploration and development of the water environment.Due to the optical characteristics of water,underwater images commonly suffer from color casts,blurriness,and other issues,leading to severe degradation in image quality.Degraded underwater images can adversely affect the visual tasks of underwater vehicles.Therefore,obtaining high-quality underwater images is crucial.Based on the principles of underwater visual imaging and depth learning algorithms,this paper has conducted the following research:(1)To address the difficulty of obtaining the true and clear features of existing underwater images and the unstable quality of paired underwater image datasets,this paper uses a self-supervised learning method based on CycleGAN to directly learn the mapping relationship between different quality image domains using non-paired underwater data;In order to solve the problem of existing enhancement algorithms ignore enhancing the subjective perception of images,this paper designs an aesthetic loss function;According to the research findings that using NIMA alone for underwater image enhancement can cause color deepening,this paper designs a histogram consistency loss;To address the problem of image detail loss that often occurs in GAN-based enhancement results,this paper constructs a two-stage generative network to enhance image details using the original image enhancement network.Comparative experiments and ablation experiments show that this algorithm has superiority in both subjective perception and aesthetics of the enhanced images.(2)Currently,most non-paired underwater image enhancement algorithms are based on the CycleGAN framework,which is complex,and existing algorithms focus more on performance,resulting in increasingly complex network structures.Therefore,to address the problem of designing complex bidirectional auxiliary networks in CycleGAN,this paper uses a one-way image domain transformation model CUT to complete the underwater image enhancement task,constructing a fast non paired underwater image enhancement algorithm;In terms of network structure,a simplified channel attention mechanism is used in combination with the UNet network,and a gating unit is used to replace the activation layer to improve image processing speed;To address the problem of too many loss functions and hyperparameters in CycleGAN,this paper designs contrastive learning loss functions and consistency loss functions based on the characteristics of underwater image datasets.In the experimental section,multiple sets of comparative experiments have been conducted,proving that the algorithm in this paper has a faster image processing speed,good enhancement results,and higher feasibility in engineering.