Research On Underwater Image Enhancement Method Based On Generative Adversarial Networ

For the past few years,with the advancement of technology,the exploration of marine resources has become increasingly frequent.The oceans are rich in mineral resources that can meet human needs.As a carrier of underwater environmental information,underwater images are an important tool for human beings to understand and explore the ocean.However,due to the absorption and scattering effects of light in the underwater environment,images captured in real underwater scenes often suffer from image blur and color distortion.It leads to great challenges for underwater imaging equipment and limits the fulfillment of underwater vision tasks.Most of the current underwater image enhancement methods are not ideal,and there exist haze residue,blurred details and unnatural colors in the enhanced results.To address this problem,this paper proposes a novel underwater image enhancement method using a deep learning approach-Generative Adversarial Network(GAN)based on the characteristics of underwater image imaging.The main content and innovation points are as follows:(1)This paper proposes an end-to-end generator,and the design of the generator is based on the U-Net encoder-decoder structure.Aiming at the problem that underwater images have different information content in different color channels and different degrees of ambiguity at different pixel positions,this paper proposes the Feature Attention Residual(FAR)module by combining the channel attention mechanism,the pixel attention mechanism,and the residual network.The FAR module enables the network to adaptively extract important features,focusing on the blurred areas of underwater images and the red channel with less information content.In order to better fuse multi-level features together,this paper proposes an Adaptive Dense Feature Fusion(ADFF)module,which can adaptively learn the spatial importance weights of different-level features,thereby enabling the network to learn from previous and current features,and learning more effective features for fusion.In order to reshape and refine the extracted features,this paper proposes the SOS(Strengthen Operate Subtract)module,which realizes the gradual refinement of image features by combining the SOS Boosting strategy.Different from other methods that only use U-Net as the generator,this paper uses the FAR module for feature extraction,the ADFF module for feature fusion,and the SOS module for feature refinement in the generator,so as to improve the learning capability of the generator.(2)The design of the discriminator in this paper is based on Markovian discriminator(Patch GAN).Different from other methods using Patch GAN as the discriminator,in order to make the discriminator better supervise the learning process of the generator,this paper designs a Multi-scale Feature(MF)module in the discriminator.This module can aggregate multi-scale features with different receptive fields,enabling the discriminator to use the feature information of global semantics and local details to achieve a more accurate discrimination process,and guiding the generator to generate more detailed and clear images.(3)This paper constructs a novel loss function that is beneficial for GAN training.In order to stabilize the training of GAN,this paper combines the adversarial loss of CGAN(Conditional GAN)and WGAN-GP(Wasserstein GAN with Gradient Penalty),and introduces the combined loss as the overall loss function of this method.The novel loss function can stabilize the training of the network,better constrain the network model,and promote the network to obtain high-quality and clear images.The experimental results show that the method in this paper can effectively solve the fogging blur and color cast problems of underwater images,and generate underwater images with clear visual effects.Meanwhile,compared with several other underwater image enhancement methods,this method achieves the best performance in both qualitative and quantitative comparisons.Among them,on synthetic datasets,compared with the latest deep learning underwater image enhancement method UIE-DAL,the structural similarity index(SSIM)of this method is improved by 6.7%,and the peak signal-to-noise ratio(PSNR)is improved by 37.8%.