Underwater Image Enhancement Based On Improved U-Net Model

In recent years,with the gradual scarcity of terrestrial resources,the development and utilization of marine resources have gradually become the strategic focus of all countries in the world.In the fields of marine military,underwater biological identification and detection,underwater robot,etc.,the acquisition of underwater images is a prerequisite for obtaining underwater information,however,due to the complexity of the marine environment and the special nature of the underwater environment,the formation of underwater images is affected by the absorption and scattering of light in water,resulting in the acquired underwater images presenting a variety of problems such as low contrast,blurred image details,color shift,and unclearness.Therefore,it is of great research importance to enhance the quality of underwater images.Classical underwater image enhancement algorithms have limited enhancement effect and enhancement efficiency for underwater images in different water quality,and cannot be better applied to underwater images in different scenes due to the lack of a priori information of water quality in different waters.Due to the excellent performance of deep learning in the field of computer vision,many researchers use deep learning methods to enhance underwater images by extracting underwater image features,which improves the effect and efficiency of enhanced images to a certain extent.However,there are still two key issues.On the one hand is the difficulty of obtaining underwater data is a current difficulty because deep learning requires a large dataset,which is difficult to obtain in deep sea environments.On the other hand,the existing deep learning methods have good enhancement effect on underwater images in some specific scenes,but it cannot fully extract underwater image features,and cannot be widely applied to real underwater images in different scenes.In this paper,we address the above two problems:1)To address the current problem that deep learning requires a large number of paired training image sets,a new underwater imaging model is proposed in this paper.Generative adversarial network is used to simulate the process of underwater imaging model,in which the generator simulates the process of forming images through absorption,scattering and atomization of light.The discriminator guides the parameters of the underwater imaging model to learn adaptively through the real underwater images,which solves the shortage of manual parameter design based on the underwater imaging model,and makes the image generated by the generator consistent with the style of the underwater image,thus obtaining the underwater data sets of two styles.2)To address the problem that the existing underwater image enhancement methods cannot fully extract the features of the underwater image,which leads to the insufficient versatility of the model,this paper uses the U-Net network with excellent feature extraction ability as the basic network for underwater image enhancement to fully extract the features of the underwater image,Which can effectively enhance the underwater images in different scenes,and at the same time,introduce the attention mechanism module to improve the contrast between the target area and the background area of the underwater image,improve the overall quality of the underwater image,and finally achieve the purpose of underwater image enhancement.In this paper,the proposed algorithm is compared with other representative enhancement methods on three different types of datasets.The experimental results show that the proposed algorithm can correct the color of underwater images in different scenes,improve the clarity and contrast of the images,and make the images conform to human visual perception,which is superior to other comparison algorithms.