The noise in an image having text is removed by convolving a shaped kernel centered on a pixel for each pixel in the image. The shaped kernel has a shape configured to identify pixels that are not part of the text. For example, the shaped kernel may be shaped with zeros in a center of the kernel to identify pixels that are not part of the text. A value for the pixel is set to erase the pixel when the resulting convolution value for the pixel is less than a threshold. The process may be repeated multiple times for differently shaped kernels, including kernels of different sizes and different configurations, such as having values greater than one in at least one of a row, column, and diagonal.