Image binarization methods and apparatus are described. A set of input image pixel values, e.g., a set of grayscale values corresponding to an input image, is processed to determine whether to recommend to use local binarization thresholds or a global binarization threshold. Edges including edge pixels are identified. A first histogram corresponding to edge pixel values and a second histogram corresponding to image pixel values are generated, subjected to one or more smoothing operations, and truncated, based on information derived from the edge histogram. Characteristics of the histograms including, e.g., minima, maxima, points of inflection, and hidden peaks, are determined, evaluated, and used to decide between local binarization thresholds and a global threshold. Based on the recommendation, a global threshold is used or local thresholds are used to process the set of input image pixel values and generate a corresponding set of bi-level values.