Research On The Restoration Method Of Water Degraded Image Based On Deep Learning

With the development of technology,research and detection of the underwater environment has become the most important research tasks in the world.Analysis of underwater images is of great significant to carry out these tasks,and the clarity of these images affect the final research results.However,due to the absorption and scattering of suspended particles,the objects underwater are difficult to identify and exhibit varying degrees of blue-green color deviation,with blurred boundaries and low contrast,which cannot reflect real underwater scene.Therefore,it is challenging and valuable to carry out research on the correction and restoration of underwater distorted images.This paper studies underwater image restoration algorithms.In view of the limitations of existing traditional algorithms and algorithms based on deep learning,such as low generality of algorithms,inaccurate color cast correction,inaccurate reconstruction of details in the image and lack of information in real images in training process,this paper designed and proposed an end-to-end deep learning underwater image restoration network to restore water degraded images,remove the effects of color shift,and restore clear boundaries.The innovative work of this article can be summarized as follows:?.Research on multi-scale underwater image restoration algorithm based on codec unit.Aiming at the limitations of existing underwater image processing algorithms in network design concept,training data and network structure,a new underwater image restoration network is proposed.The network adopts an end-to-end structure,which directly avoids the error superimposed problem.Meanwhile,this paper proposes a new codec unit,with the combination of multi-scale idea and the usage of cross-layer connection and codec structure,the ability of the algorithm to analyze and reconstruct underwater images can be enhanced.Images recovered by this network have higher contrast,more natural colors,and higher overall image quality than those recovered by other algorithms.When dealing with severely degraded underwater images,the network can obtain significantly better results than other algorithms.?.Research on underwater image restoration algorithm aided by ambient light perception based on dense connections.Existing algorithms have areas for improvement in how to efficiently learn features between images before and after degradation,and how to accurately correct colors.This paper proposes a new underwater image restoration network based on these improvements.The network introduces the ambient light perception sub-network,so that the results obtained by network restoration have a more natural color presentation than other methods.Simultaneously,the network introduces a residual module under dense connections,which makes the information transmission of each layer of the network more efficient during the training process,and finally obtains more complete feature information.Through subjective and objective experimental comparison,the image processed by this algorithm has clearer boundaries,more natural colors,and higher overall image quality than other methods.?.Research on alternate training method of color compensation.Considering that the existing deep learning algorithms lack real underwater image information in the training process,this paper proposes a color alternate compensation training method that can reasonably use real underwater image information.The underwater images with color swatch and the corresponding ideal color value are applied to make a real underwater image training set,which provides real color information in order to compensate for the deviation of the synthetic image and the real situation in this respect.Experiments show that the network trained with this method can have a more accurate presentation in image color restoration than the network trained without this method.