Research On Underwater Image Enhancement Using A Deep Adversarial Network

Clear underwater images play a vital role in the information obtained by the ocean.However,due to the underwater medium and abundant suspended particles,light is absorbed and scattered in the underwater scene.As a result,the raw underwater image suffers from the serious degradation,such as color distortion,low contrast,and fuzz.To obtain the clear underwater scene,it is great significance and valuable to adopt underwater image enhancement method.To eliminate the fuzz effect and color distortion of the underwater scene,we design the underwater image enhancement method based on fusion dark channel prior and color correction.First,the proposed method obtains the dehazed image by the fusion dehazing method,which is calculated on the difference between maximum green-blue dark channels and maximum red dark channel,and the top 0.1%bright pixels in green-blue dark channels.Second,the proposed method obtains the color corrected image via the color restoration method corresponding to the human visual system.Finally,the proposed method employs an efficient and simple weight fusion strategy to incorporate the dehazed image and the enhanced image for yielding the final high quality underwater image.Moreover,the proposed method can improve the other degraded image quality,such as low-light image and haze image.To increase the generalization of the final results,we design the underwater image enhancement method using the multiscale dense generative adversarial network.The effective combination of residual learning,dense concatenation,and multi-scale can improve network performance,render more details,and utilize previous layer image features,respectively.The spectral normalization stabilizes the training of the discriminator.The meaningful adversarial loss including the ₁L loss and gradient loss is adopted to preserve image features of ground truth.Extensive experiments are provided to demonstrate the superiority of the proposed method.Furthermore,the ablation study is carried out to show the contributions of each component.To further explore the impact of underwater image quality on vision tasks,our paper employs the classic and widespread used application tests,such as keypoint matching,edge detection,and object detection.Extensive results of application tests indicate that the existing underwater image enhancement methods can effectively improve the performance of low-level vision tasks,but cannot improve the performance of object classification of high-level vision tasks.