The document discusses a proposed unified framework for progressively recovering images from corrupted observations through hybrid graph Laplacian regularization. The framework first constructs a multiscale representation of the target image using a Laplacian pyramid. It then progressively recovers the degraded image from coarse to fine scales, allowing for recovery of sharp edges and texture. Within each scale, a graph Laplacian regularization model is learned that minimizes error while preserving image structure. Between scales, the model is extended to a high-dimensional feature space to describe interscale correlations and propagate local structure regularity from coarser to finer scales. Experimental results on impulse noise removal demonstrate better performance than state-of-the-art algorithms.