Underwater Image Enhancement Via Aggregated Physical Prior And Deep Networks

Underwater image enhancement plays an important role in practical applications such as underwater target detection and bottom-fishing.However,underwater degraded images are often affected by low visibility,low contrast and color distortion.When dealing with the above degradation problems,pixel-based image enhancement methods have certain limitations in removing fog effects,due to the discard of physical imaging model;model-based methods may fail in some certain areas,and sometimes may introduce artificial noise and artifacts during the scene restoration;because training data can hardly cover various scenes and the lack of effective physical domain knowledge,the performance of data-driven deep learning methods may be limited by the training set and make it difficult to be used generally.In view of above degradations problems and the deficiencies of the existing methods,based on the improved imaging model,this paper integrates data-driven deep residual convolutional neural networks(CNNs)into the propagation scheme,to jointly estimate the transmission map and restore latent clear scenes.Considering the possibility of artificial light sources and the possible extra overexposure introduced in restoration,firstly we introduce the scene residual term in imaging model,and then the half-quadratic algorithm is used to minimize the energy model which contains the implicit prior of regularization term.Guided by physical domain knowledge,we design a framework that jointly optimizes the transmission and restores clear scenes iteratively,then the undesired artifacts can be suppressed in the restoration process.In each iteration,the deep residual network was integrated to estimate regularization terms of the transmission map,clear scenes and scene residuals.Hence,the proposed framework can benefit from the advantages of data-driven networks such as flexibility and computational power.In addition,different from end-to-end deep CNN models,as the purpose of the CNNs plugged in the proposed framework is to learn priors of regularization terms,it is unnecessary to synthesize specific training data for different underwater scenes.Therefore,the proposed framework can be more adaptable with lower training costs.In addition,this paper designs a novel residual architecture to aggregate a priorand data information to estimate underwater transmittance.This architecture bridges the differences between previous drive models and data-driven networks,taking advantage of its advantages but avoiding the limitations of prior models.And a lightweight learning framework is proposed to train the transmission network of transmittance.In order to more comprehensively evaluate and verify the effectiveness of the proposed method,this paper builds a large-scale real-world underwater image database that can be used to evaluate algorithms from multiple perspective of improving image quality,correcting color shift and task-driven performance.Both experimental qualitative and quantitative analysis show that the proposed method can effectively suppress artifacts and local overexposure while removing the underwater fog effect and correcting color shift,and achieve the best numerical results on various evaluation indicators.It fully proves that the combination of traditional physical prior and deep CNNs can effectively solve image restoration and image enhancement problems.