Research On Underwater Image Sharpness Methods Based On Generative Adversarial Networks

Underwater images play an important role in obtaining Marine information,but due to the different attenuation speeds of light at different wavelengths in the water,underwater images will have color distortion and insufficient underwater light intensity will cause low contrast in underwater images,which seriously affects the further use of underwater images,so it is necessary to correct the color shift,enhance the contrast of the images for the clear underwater images.The underwater images sharpness methods based on generative adversarial networks train the network in a data-driven manner,which can deal with a variety of underwater image degradation problems and has good robustness.Based on the generative adversarial networks,this thesis studies the underwater image sharpness problems.The main work is as follows:The conditional generative adversarial network is used to study the sharpness of underwater images.Based on StarGAN,underwater images are classified by adding a discriminative model,so that the network can adopt different enhancement strategies according to the different attenuation degrees of images.The residual module is introduced into the model to increase the network depth and enhance the network’s ability to extract high-level semantic information of the image.In addition,by optimizing the loss functions,the constraint on the network output result is realized,so that the content and structure of the processed images and the input images are consistent.The designed underwater image processed by the network has been significantly improved in terms of color and sharpness.Muti-scale generative adversarial network is used to study the sharpness of underwater images,and a new network is constructed.Multi-scale generative model is used to extract rich image features and improve the network’s feature expression ability.Then,multi-scale discriminative model is used to further supervise the training of the generative model to make the high-resolution underwater images output by the whole network clearer.In addition,the parameters of the network are optimized through the weight sharing strategy,which is more conducive to the migration and reuse of the model.