Visibility Restoration of Single Haze Images Based On Feature Aggregation

In order to defect drawbacks of Dark Channel Prior(DCP)-based dehazing methods, which suffer from undesirable artifacts and color distortion, we introduce multidimensional theory to achieve effective prediction of haze free images. Local manifold similarity is first employed to reduce the initial estimation error. And then we employ signal filtering method based on contrast-based Gaussian curvature to refine the rough transmission map. The experimental results have shown that the images restored by our algorithm are more clear and natural than start-of-art approaches.