去雨算法的PYTHON完成,Rainstreakscanseverelydegradethevisibility,whichcausesmanycurrentcomputervisionalgorithmsfailtowork.Soitisnecessarytoremovetherainfromimages.Weproposeanoveldeepnetworkarchitecturebasedondeepconvolutionalandrecurrentneuralnetworksforsingleimagederaining.Ascontextualinformationisveryimportantforrainremoval,wefirstadoptthedilatedconvolutionalneuralnetworktoacquirelargereceptivefield.Tobetterfittherainremovaltask,wealsomodifythenetwork.Inheavyrain,rainstreakshavevariousdirectionsandshapes,whichcanberegardedastheaccumulationofmultiplerainstreaklayers.Weassigndifferentalpha-valuestovariousrainstreaklayersaccordingtotheintensityandtransparencybyincorporatingthesqueeze-and-excitationblock.Sincerainstreaklayersoverlapwitheachother,itisnoteasytoremovetheraininonestage.Sowefurtherdecomposetherainremovalintomultiplestages.Recurrentneuralnetworkisincorporatedtopreservetheusefulinformationinpreviousstagesandbenefittherainremovalinlaterstages.Weconductextensiveexperimentsonbothsyntheticandreal-worlddatasets.Ourproposedmethodoutperformsthestate-of-the-artapproachesunderallevaluationmetrics.
1