In this work, a new clustering based binarization technique has been proposed. Clustering is done depending on the information density of the input image. Here input image is considered as a set of text, images as foreground and some random noises, marks of ink, spots of oil, etc. in the background. It is often quite difficult to separate the foreground from the background based on existing binarization technique. The existing methods offer good result if the input image contains only text. Experimental results indicate that this method is particularly good for degraded text document containing graphic images as well. USC-SIPI database is used for testing phase. It is compared with iterative partitioning, Otsu’s method for seven different metrics.