In this paper, we have addressed a new approach to encode the DCT coefficients efficiently by exploiting the structural differences among the probability distribution functions of them. We have proved analytically that if signals are separated according to their source models, the entropy of the separated signals decreases below that of the unseparated signals. And simulation results show that the average codelength decrease when each (runlength, level) symbol is encoded adaptively according to the index associated with the symbol.