Single Image Reflection Removal With Feature Exclusion And Compensation

When people take pictures through translucent media such as glass,the camera will capture both the target scene and the reflected scene,and get degraded image with reflection artifacts.The main purpose of image reflection removal is to recover the clean background image from the degraded image with reflection.With the development of computer and artificial intelligence technology,convolutional neural network has been widely used in computer vision tasks and achieved remarkable results.The existing methods lack explicit constraints on the feature maps in intermediate layers of the network,and solely apply constraints on the final estimated results to make the network get limited ability of reflection separation.This thesis proposes a multiscale single image reflection removal method with feature exclusion.Taking the feature space as the starting point,this method maximizes the distance between the features of the two components that are to be estimated in the feature space,and further reduces the coupling of information between the features of the two components.In this way,the method achieves better reflection separation results.Some deep learning based methods of image reflection removal adopt a dualbranch network structure to restore the background layer and reflection layer in parallel.However,the two branches remain independent in the inference process and lack the information interaction in stages.It makes the two branches not able to obtain the current information of each other as supplement and guidance.This thesis proposes a residual-based feature interaction and compensation mechanism.It makes the two branches of the network can transfer information to each other stage by stage through the residual to guide and complement each other,so that the two branches can promote and enhance the recovery results of the two components of the background and reflection images.In order to solve the problem of over-optimization of the feature exclusion mechanism mentioned above,this thesis also improves it and proposes a slack feature exclusion mechanism.In addition,in order to avoid the loss of information caused by feature misalignment,a feature alignment strategy is also proposed to reduce the discrepancy between the internal patterns of the background and reflection features at the same stage.Based on the feature interaction and compensation mechanism,slack feature exclusion mechanism and feature alignment strategy,this thesis also proposes a single image reflection removal method with slack feature exclusion and compensation.From two aspects of objective evaluation metrics and subjective visual quality,this thesis compares the two proposed methods with several state-of-the-art methods on four benchmark datasets,and proves the superior performance of the proposed methods.The ablation study also proves the effectiveness of the main innovations proposed in this thesis.