The spatial and temporal resolution of surface grain-size characterization is constrained by the limitations of traditional measurement techniques. In this paper we present an extremely rapid image-processing-based procedure for the measurement of exposed fluvial gravels and other coarse-grained sediments, defining the steps required to minimize the errors in the derived grain-size distribution. This procedure differs significantly from those used previously. It is based around a robust object-detection algorithm that produces excellent results on images exhibiting a wide range of sedimentary conditions, crucially, without any user intervention or site-specific parameterization. The procedure is tested using a data set comprising 39 images from three rivers with contrasting grain lithology, shape, roundness, and packing configuration and representing a very wide range of textures. It is shown to perform more consistently than the best existing automated method, achieving a precision equivalent to that obtainable by Wolman sampling, but taking between one sixth and one twentieth of the time. The error in area-by-number grain-size distribution percentiles is typically less than 0.05 ψ.