Research On Underwater Image Enhancement Based On Deep Learning

As human exploration gradually touches the ocean,the demand for underwater images in various fields is also increasing.However,affected by various factors in the water,such as impurities,light scattering,etc,it is often difficult to obtain clean images to satisfied what they want actually.As deep learning has a rapid growth,it has excellent performance and exquisite structure,which made a significant breakthrough in the field of underwater image and fully solved the problem of underwater image degradation.Hereto,this study conducted research on underwater image enhancement tasks based on deep learning methods,and its key contributions are as follows:(1)In view of the fact that traditional convolutional neural networks generally only focus on local input information,rather than comprehensive attention,and are unable to extract sufficient semantic information,a feature pyramid and convolutional attention fusion network is introduced at length in this paper.The model adopts encoder-decoder architecture as a whole.In the encoder,it is divided into four stages to extract semantic information by gradually raising the level.In the decoder,FPN was referred to and combined with skip connections to construct a feature pyramid structure.The semantic information was fused from top to bottom after each stage of the encoder,combining low semantic and high semantic information to enhance the model’s overall ability.In addition,SE module is also introduced in feature extraction of each stage to improve the model’s semantic extraction performance extensively.The experiments on the dataset UIEB showed that the model achieved 24.33,0.92 and 3.39 on three objective evaluation indicators of underwater images,PSNR,SSIM,and UIQM,which basically exceeded the six models used for comparison,such as Deep-SESR,FUn IE-GAN,Shallow-UWnet,etc,and effectively solved various problems in underwater images.At the same time,the ablation experiments on the dataset UIEB,EUVP and UFO-120 proved the effectiveness of the added attention mechanism SE module.(2)For the problem that the traditional Non-local module cannot fully utilize information during the global information adoption process,resulting in insufficient improvement of model performance,an improved Non-local module is mooted.Unlike the traditional Non-local module,this module has been extended,the number of model layers has changed from one to three and the convolutional kernel size is designed as 2,4 and 8.By adding the layers,multi-scale semantic information is obtained.The experiments on UIEB dateset show that the improved Non-local module can make more full use of global information than the traditional Non-local module,thus improving the overall performance of the model.(3)Aiming at the common phenomenon of semantic information loss in generative adversarial networks,a dual attention fusion generative adversarial network is proposed and its whole structure adopts Pix2 Pix.In the generator,the dual attention mechanism UNet network is designed according to DANet,the improved Non-local mechanism is introduced in the head,and the Transformer Encoder module is introduced in the tail,thus reducing the semantic loss in model training.In the discriminator,it is modeled as Patch GAN.By enhancing the generator’s ability to synthesize images,it can improve the discriminator’s ability to recognize true and false images in the confrontation training,and ultimately improve the comprehensive ability of the model.In the experiment on the dataset EUVP,the three objective evaluation indicators of underwater images,PSNR,SSIM and UIQM of the model reached 25.07,0.86 and 3.45,respectively,while the corresponding indicators on the dataset UFO-120 reached 27.04,0.88 and 3.39.Experiments have shown that the performance of the dual attention fusion generative adversarial network exceeds that of the latest deep learning model as a comparison,thus better solving the problem of underwater image degradation.