Study Of Restoration Technology On Underwater Turbulent Distorted Image Based On Deep Learning

Image is one of the important ways to convey information,it can offer information to humans in an intuitive way.However,images may be subject to various disturbances during shooting,resulting in loss of information.When shooting underwater scenes through the water surface with camera device,the turbulence of the water surface may cause the captured image geometric distorted.Due to the important guiding significance of clear underwater images for industrial production,many countries have applied research on the recovery of underwater distorted images.However,China is still in its infancy in this respect.Most of the previous underwater distortion image restoration techniques are based on video sequence.With the rise of deep learning,it has become possible to recover from single-frame underwater distorted images.However,current method is still not ideal for imaging quality.This paper has carried out related research on the recovery technology of underwater turbulence images.The main works are as follows:1.An underwater turbulence distorted image restoration algorithm based on dense connection is proposed.The underwater distortion image is restored by using densely connected generative adversarial networks,and the Wasserstein generative adversarial networksgradient penalty(WGAN-DP)loss function is introduced to stable the training.The Structural Similarity(SSIM)and Peak Signal to Noise Ratio(PSNR)on the validation set reached 0.5595 and 19.925,respectively,and the visual effects is improved;2.An underwater turbulent distorted image restoration algorithm based on attention mechanism is proposed.By introducing attention mechanism,a two-stage image restoration algorithm framework is designed.Firstly,an attentive distortion network is used to learn the distortion field of underwater distortion image and generate an attention map;then perform a preliminary distortion recovery on the input image according to the distortion field,and input the concatenation of output image and the attention map into another network to recover the details,further improving the image quality of the image.The SSIM and PSNR on the validation set reached 0.6022 and 20.522 respectively,and the visual effect is also significantly improved.