Research On End-to-End Underwater Image Enhancement Algorithm Based On Bayesian Theory And Deep Learning

In recent years,on the one hand,due to the excessive consumption of land resources,and on the other hand,due to disputes over maritime territories,people have turned their attention to the ocean,which has an area more than twice that of land and more abundant resource reserves.Due to the collection and application of underwater environmental information,including but not limited to underwater military applications,autonomous underwater vehicles(AUVs),seabed terrain and geomorphology exploration,the search and recognition technology of underwater targets plays an extremely important role.Underwater images are an important carrier that can intuitively transmit underwater environmental information,with higher resolution than sonar systems and the ability to process color information.Therefore,they play an irreplaceable role in the field of underwater target recognition.However,seawater itself has different absorption rates for light of different wavelengths,which can cause serious color distortion in underwater images;The soluble inorganic salt,inorganic compound matter and organic matter particles contained in seawater will seriously absorb and scatter the underwater light,resulting in blurred image edges and reduced contrast.Therefore,in the face of underwater images that are difficult to extract image feature information,it is necessary to preprocess the underwater images through image enhancement technology to improve the perceptual quality of the images,so that they can be adopted in advanced computer vision tasks.This study proposes an underwater image enhancement algorithm based on Bayesian retina theory,combining the advantages of various prior knowledge.The algorithm uses a simple and effective color correction method based on statistics to solve the problem of color distortion of underwater images,and then establishes a Bayesian based probability model of retina to solve the problem of insufficient exposure and blurring of images.A maximum a posteriori(MAP)formula for retinal underwater image enhancement was established using a multi step degree prior on color corrected images.First and second order gradient priors were applied to reflectance and illumination to capture finer and more complete structures from underwater images.Subsequently,a complex underwater image enhancement problem was transformed into two simple denoising problems,and their convergence was mathematically proven and analyzed.Finally,experiments are carried out to demonstrate the effectiveness of the proposed method in color correction and naturalness preservation,structure and detail enhancement,and artifact or noise suppression.Deep learning is one of the hot spots in underwater image enhancement research.In order to meet the need of lightweight deployment of deep learning network models in underwater terminals,this study proposes an improved shallow end-to-end underwater image enhancement model based on Shallow UWnet model.The model consists of a series of convolution blocks,batch normalization layers and Leaky Re LU activation function,and uses a method that includes mean square error loss The joint loss function of perceptual loss and structural similarity loss.The experimental results demonstrate that the performance of this network is no less than that of the mainstream deep neural network underwater image enhancement methods,with smaller volume and faster processing speed.Finally,this article compares the performance of two algorithms in image clarity and concludes that in terms of removing fog and improving image brightness,and in terms of improving image contrast,different image processing methods should be selected according to different needs to achieve the optimal effect.