Super Resolution Underwater Image Enhancement Algorithm Based On Color Correction

The ocean contains resources such as metals,natural gas and oil,and is one of the most popular exploration areas at this stage.Collecting marine information by artificial methods not only causes great harm to diving operators,but also has extremely low work efficiency.Underwater robots have become an important means of ocean exploration due to their flexibility,large diving depth and multi-functionality.However,the marine environment is relatively complex.Affected by the water medium and suspended particles,natural light entering the water will absorb and scatter,so that the images captured by robots often have serious color deviation and fog.At the same time,due to the influence of ocean terrain and the limitation of shooting equipment,underwater images still have problems such as blurred details and low resolution.These problems make it difficult for the underwater robot to obtain the required information in the image,which affects the recognition and grasping of the target by the robot,and further increases the difficulty of underwater operation.Therefore,research on underwater image enhancement(color correction,dehazing)and super-resolution(restoring image details)methods has high theoretical significance and application value.However,there are few methods that can take care of both tasks at the same time.To this end,this thesis proposes a generative adversarial network model that can simultaneously achieve image enhancement and super-resolution.In addition,considering the slow image processing speed of large-scale networks and the difficulty of deploying on mobile platforms,this thesis proposes another lightweight image enhancement and super-resolution network at the same time.The research content of this thesis mainly includes the following three points:(1)This thesis proposes a method for making underwater datasets,which alleviates the problem of difficult acquisition of underwater images.In this thesis,two methods are used to obtain the original image of the ocean and the clear image corresponding to the same scene.The first method: firstly,the underwater image is captured from the diving video,and then the enhanced image(clear image)is obtained by using the traditional image enhancement algorithm.The second method: firstly,the indoor image(clear image)is taken,and then the underwater imaging model is used to add special effects to the indoor image to simulate the underwater image.(2)This thesis proposes a generative adversarial network model that can simultaneously achieve underwater image enhancement and super-resolution.Using this model,the underwater image can be color corrected,the fogging effect can be removed,and the details of the image can be recovered at the same time.The generator of this model is composed of image fusion module,feature extraction module and image enhancement super-resolution module,and the discriminator adopts Patch GAN structure.(3)This thesis presents a lightweight image enhancement and super-resolution model.To solve the problems of difficult deployment of large-scale neural network model and slow data processing speed,another lightweight image enhancement and super-resolution network is proposed in this thesis.The network model has the advantages of smaller volume,less resources,easy deployment and fast computing speed,and can be used to process underwater video data in real time.Experiments show that the underwater image processed by the generative adversarial network model proposed in this thesis is superior to other methods in visual effect.For the lightweight network,the processing time is only 60% of that of the generative adversarial network model on the premise of ensuring the image quality,which has a speed advantage compared with other methods.