Underwater Image Processing Algorithm Research Based On Imaging Model And Deep Convolution

At present,with the rapid development of artificial intelligence,underwater robots can effortlessly capture underwater images.Due to light commonly suffers from absorption and scattering as it propagates through water,the acquired underwater images usually exist color casts,low contrast,and noise.The degraded underwater image limits its application in image recognition,target detection and object tracking.The enhanced underwater image can improve the image’s visual quality and help you better grasp the underwater world.This paper improves the visual quality of underwater images and makes the processed underwater images more in line with human visual characteristics and the needs of other visual tasks from the standpoint of underwater image imaging model and depth convolutional network.This paper aims to improve the visual quality of underwater image from the perspective of underwater image formation model and deep networks.The main research contents are as follows:1)The underwater image exits some problems including low contrast,color cast and blurred details.A multi-input fusion adversarial network(MFGAN)is proposed for underwater image enhancement.Firstly,the degraded image is preprocessed in two ways: color correction and contrast enhancement.Secondly,the generated network is utilized to learn the difference between the two preprocessing images and the degraded images,which obtains image’s confidence map.Then,a texture extraction unit is designed to extract the texture features of the two preprocessing images,and the extracted texture features are fused with the confidence map counterpart to eliminate the artifacts and detail blur generated by the two enhancement methods.Finally,by constructing multiple loss functions to improve network performance.The experimental results show that the enhanced underwater image has bright color and improved contrast,the average value of UCIQE and NIQE is 0.6399 and 3.7273,respectively.Compared with other algorithms,the algorithm has significant advantages and proves its good effect.2)MFGAN,which is not an end-to-end depth network theoretically,uses traditional methods to process the image’s color and contrast.This paper proposes an adaptive learning attention network(LANet)for underwater image contrast enhancement and color correction as a solution to this challenge.Firstly,a multi-scale fusion module is proposed for fusing various spatial information.Secondly,a new parallel attention module is designed,which integrates pixel and channel attention and pays more attention to illumination features as well as more significant color information.The adaptive learning module may maintain the shallow information as well as learn the key feature information adaptively.In addition,a polynomial loss function is constructed,which is composed of average absolute error and perceptual loss.Finally,an asynchronous training mode is introduced to improve the network performance.Qualitative analysis and quantitative evaluation show that this method has good performance on different underwater data sets.3)MFGAN and LANet have more convolution layers,which means they have more parameters,which affects network processing speed.Thus,a physical model and comparison learning(PCNet)is proposed for fast underwater image restoration.Firstly,the correation network is proposed,including two functions.One provides a middle image to correct the color casts of the reconstructed image by underwater image formation model,while the other generates feature information using the number channel 32 as the parameter estimation network’s input.A parameter estimation network with two branches is designed to accurately estimate background light and transmission map.In addition,contrastive regularization is proposed to generate visually clearer underwater images.The values of Parameters and FLOPs of PCNet are 0.17 M and 5.63 G respectively.Compared with other algorithms,PCNet has lightweight and effectively improves the visual quality of underwater images.The paper has 46 figures,24 tables and 61 references.