Research On Underwater Image Enhancement Algorithm Based On Deep Learning

Nowadays,marine scientific research has increasingly relied on underwater images taken by autonomous underwater vehicles(AUVs)or remotely operated underwater vehicles(ROVs).Natural light will be absorbed and scattered when propagating underwater,resulting in serious visibility problems in underwater images,specifically in terms of color shift,lack of contrast and low clarity.Severe degradation not only affects the quality of underwater images,but also limits the progress of underwater vision.At present,underwater image processing methods mainly include traditional methods and deep learning-based methods.Traditional methods have poor generalization in different underwater environments due to complex physical and optical factors underwater;deep learning-based methods do not perform well in underwater image details because of network structure.This paper focuses on the imaging characteristics of underwater images and mainly uses deep learning methods to improve the image quality.Our innovations are as follows.(1)In order to recover underwater image details better,we propose a generative adversarial network based on dense feature fusion module and deep boosted module,which is capable of enhancing underwater images end-to-end.In order to enhance the spatial information extraction capability of the generator,a dense feature fusion module is designed using the back-projection mechanism,which is able to utilize non-adjacent network layer features,thus preserving the missing spatial information in the high-resolution features.Meanwhile,the deep boosted module is used to combine the features of the corresponding layers to gradually recover the image details,which makes the network retain more detailed information.Secondly,in order to solve the problems of gradient disappearance and pattern collapse in generative adversarial networks,we introduce WGAN with gradient penalty to optimize the training process of generators and discriminators.WGAN can improve the diversity of generated samples.Finally,algorithm of this paper has the highest score in each image evaluation index when compared to other methods.The ablation experiments also show the effectiveness of the proposed dense feature fusion module and deep boosted module in detail feature recovery.(2)To further solve the problem of real-time application of underwater image enhancement algorithms,this paper proposes a super-resolution-based underwater image enhancement model,which not only can recover the color of underwater images,but also can improve the resolution of images at a scale of 2x and 4x.It brings convenience for real-time application.Firstly,we introduce the residual in residual dense block,which is a structure that can improve the feature retention ability of the model and make the training of the model more stable.In addition,the model introduces the attention mechanism that helps the network to learn the foreground regions of the image,focusing on the foreground regions of the image to enhance the global contrast of the image.Finally,the model employs a multiscale loss function to supervise its training and optimize the color,sharpness and contrast,respectively.Our method works very well in image enhancement and super resolution when compared to other methods.The ablation experiments prove that the proposed multiscale loss function in this paper has significantly improved the objective indexes of color(UICM),sharpness(UISM),and contrast(UIcon M)of enhanced underwater images.Finally,target recognition experiments were conducted on seafood videos taken by underwater robots,and the rate of seafood recognition was greatly improved.It verifies that the super-resolution-based underwater image enhancement algorithm has good application prospects.