We present a fast, non-iterative technique for producing grayscale images from dithered and error diffused halftones. Our inverse halftoning scheme is comparable in quality to the best kernel estimation, wavelet, and MAP estimation methods in the literature, yet requires a fraction of the computational cost. When compared to the MAP estimation method, our scheme is 6 times faster for error diffused halftones, and 180 times faster for dithered halftones. The technique is a completely feedforward technique with pixel-level parallelism. It consists of one Gaussian filter and one median filter in the first stage, and two Gaussian filters, a thresholding operation, an edge map computation, and a median filter in the second stage. The goal of the second stage is to enhance the rendering of edges in the grayscale image. We compare the quality and computational requirements of our technique with the best methods described in the literature.