| The most important basis for image processing is the color information of the image.Because of the difference between the imaging device and the illumination environment,the color of the captured image may deviate from the color perceived by the actual human visual system.However,the color deviation may affect the accuracy of the image analysis result.Therefore,color correction is particularly important.Color correction is a color constancy problem.For the problem that the imaging device is sensitive to the illumination conditions and the color characteristics of the acquired images are greatly changed,the color of the image is corrected by several traditional color correction algorithms.The traditional color correction algorithms used in this thesis include the Gray World algorithm,Histogram Equalization algorithm,Retinex algorithm.The Retinex algorithm includes SSR(Single Scale Retinex),MSR(Multi Scale Retinex)and MSRCR(Multi Scale Retinex with Color Restore).In this thesis,several different traditional color correction algorithms are used to correct the foggy image and low-illumination image respectively.The corrected image is evaluated by several different objective image quality evaluation criteria,such as contrast,average gradient,information entropy and NR-IQA(No-Reference Image Quality Assessment).The results show that the MSRCR algorithm with color recovery factor is superior to other algorithms.Because the details of the image are kept the most complete,which is more consistent with the structure of the real scene,effectively solving the local color distortion caused by the image enhancement process.In the process of exploring the ocean,people need to accurately analyze the acquired underwater images and to restore the true color information of the underwater scene as much as possible.We conclude that the traditional color correction algorithm is not effective through experiments,so we use the generative adversarial network to perform color correction of underwater images.Neural networks can implement end-to-end modeling of complex nonlinear systems,but neural network structures require more underwater images and images that reflect the true color of underwater scene.But in real life,it is difficult to obtain images that reflect the true color of underwater scene.In response to this problem,the generativeadversarial network is introduced into the color correction algorithm in this thesis.Firstly,the image of the natural scene and the corresponding depth map are used as input of the generator to generate underwater image corresponding to the natural scene through three steps of attenuation,backscattering and camera model.Then the generated underwater image and the real underwater image are used as input of the discriminator,the discriminator will discriminate the real underwater image as 1 and the generated underwater image as 0.In the process of alternately training the generator and discriminator networks,the generator improves the generating ability by updating its own network weights continuously,and the discriminator improves the discriminating ability by updating its own network weights continuously.When the network reaches convergence,the discriminator cannot distinguish whether the underwater image is real or generated,indicating that the model has been trained.Secondly,the generated underwater image and the image of the natural scene are used to train the color correction network,the color correction network can be divided into two stages: relative depth estimation and color recovery.Both of these stages use a deep full convolutional neural network based on SegNet.In order to preserve the texture information in the image color restoration so that the corrected image can still be processed,a jump layer is added to the basic codec structure to compensate for the information loss caused by the high frequency information passing through the network.Because high frequency information can be transmitted from the encoder to the decoder by the jump layer,which can improve the convergence speed of network training and improve the recovery quality of underwater images.Thirdly,after the color correction network is trained,the corrected image can be output by inputting the real underwater image.We adopt an objective method of NR-IQA.The model trained in the algorithm of the generative adversarial network learns the mapping relationship from the underwater image to the image of the natural scene,so that the effect of underwater image correction is better and the color information is more realistic. |
