Research On Underwater Optical Image Quality Perception And Visibility Improving Algorithm

Underwater optical imaging has played an important role in many fields,such as marine research,marine development,and underwater engineering.However,due to the influence of light absorption,light scattering,and image compression,the acquired underwater image generally has common problems such as low sharpness,low contrast,and color deviation.In the acquisition stage of the underwater robot,the input image is perceived by a non-reference image quality evaluation method.Underwater image which can be further clarified is select to be compressed and transmitted to the backend.A two-step clearing method is utilized in the server backend to remove artifacts of the compression image and solve the attenuation and scattering problems existing in the original scene.In the stage of underwater image quality perception,this paper firstly introduces the degradation model of underwater image and proposes corresponding objective evaluation indicators according to degradation characteristics.Considering that human visual characteristic also has an important impact on image quality,this paper proposes a no-reference image quality assessment method which combines the evaluation indicators and visual attention map.Experiments show that the proposed method is more relevant to the human visual system.Wavelet transform coefficients are introduced as the prior knowledge to eliminate the uncertainty of the clear image corresponding to the compression image.This paper introduces the residual estimation network of the dual domain to estimate the residual of the pixel domain and the wavelet domain respectively.The correspondingcompression artifacts image is then generated by a generative adversarial network(GAN).An iterative residual estimation method is utilized in the pixel domain.In addition,long and short-term memory(LSTM)units are utilized to optimize the iterative estimation procedure.Compared with the existing three kinds of compression artifacts removal algorithms,the proposed method has the best performance onretaining the information of the original image while removing compression artifacts and improving the sharpness of underwater image.The output result is closest to the real image.Underwater optical image suffers various types of degradation during the acquisition process.Therefore we combine the color correction result and turbidity removal image through a multi-scale fusion framework.In order to improve the accuracy of the estimated transmission map in over-turbid waters,a convolutional neural network(CNN)is introduced to the framework.Compared with the existing methods,our method can improve the contrast of underwater image and correct the color robustly.Experiments show that the proposed method provides the best performance in terms of both subjective and objective evaluation indicators.The extended experiments show that the proposed method has a good promotion effect on feature point matching and target segmentation.