Underwater Image Restoration Based On The Deep Learning

Nowadays,with the depletion of land resources,countries all over the world are continuously expanding their exploration to the marine field.The poor underwater imaging environment leads to blur,low contrast,and color distortion in the underwater images.These degradation phenomena have seriously affected the visual quality of underwater scenes,the accuracy of image processing algorithms as well as the research fields and industrial applications,like marine ecology,underwater transportation,and marine resource exploration.Due to the extensive application of deep learning technology in the field of image processing in recent years,its excellent graphics processing capabilities and the ability to learn complex mappings have been recognized.Therefore,the underwater image restoration technology based on deep learning is a very important and meaningful research topic.The main contents of this article are as follows:First,this article refers to a large number of related papers on image restoration algorithms and deep learning technique.Introduced and analyzed the common algorithms and research status of deep learning technology and underwater image restoration technology of recent years.At the same time,the advantages and disadvantages of various image restoration algorithms are discussed and analyzed,which provides a basis for the follow-up research work.Second,in view of the problem that artificial light sources will cause many restoration algorithms to fail,the traditional underwater image formation model is revised by adding an nonhomogeneous background light term.And according to the modified formation model,a method of synthesizing underwater image datasets is proposed.Comparing with other underwater synthetic datasets,our dataset takes into account the nonhomogeneous background light phenomenon in the environment.The supervised learning model trained by our dataset has the ability to eliminate the over-enhancement and over-compensation phenomenon in the restored images.Third,a convolutional neural network architecture with decomposition and fusion structure is proposed,called FMSNet.In this network,instance normalization is used to improve the domain adaptability of the network,and the residual connection structure is used to increase the fitting ability of the network.The feature map decomposition module based on Gaussian filtering allows the network to perform image restoration tasks more efficiently.Experiments show that the training performance of this network is better than some widely used deep learning network structures.By ablation study,we prove that each component of the network has its own contribution to the performance.In addition,a lightweight version of the network,called FMSNet-B network,is proposed.Under the premise of ensuring the training performance,FMSNet-B network can reduce the number of network parameters by over 95% of the original,and the computational amount can be reduced by more than 80%.Fourth,an image restoration algorithm based on FMSNet network is proposed.By using our underwater synthetic dataset to perform supervised learning on the FMSNet network,the obtained restoration model can directly map the underwater degraded image to its restored version without any pre-processing and post-processing.Comparing with other common underwater image restoration methods,evaluation results on synthetic underwater images and real underwater images prove that our method can significantly improve image clarity and correct color information relatively accurately.Evaluation by the SIFT algorithm shows that the restoration method proposed in the thesis can increase the average number of matched feature points by 285%.