Reduced Complexity Architecture for Convolution Based Discrete Cosine Transform

Discrete Cosine Transform is a popular transform used in signal/image processing applications. Reduction in complexity of hardware architecture for the computation of DCT using the convolution based algorithm is proposed. An N point DCT can be computed through 2 pair of [(M-1)/2] point cyclic convolutions where M is an odd number such that N=2M. The proposed architecture uses only a pair of systolic array where inputs are pipelined against the one in literature where 2 pairs of systolic arrays are used. One of the systolic arrays uses processing element with tag bit and the other one does not need a tag bit. The architecture uses 50% less number of processing elements with just an additional increase in computation time by one unit. The architecture is divided into three stages as preprocessor stage, compute stage where a systolic array computes the cyclic convolution and a post processing stage to process the output of the systolic array to get the actual DCT output. It is observed that the proposed architecture has a reduction of about 42%adders and 64% multipliers as compared to the one in literature. Further, the architecture is simulated in ModelSim 6.5 and synthesized using Xilinx ISE10.1using Vertex 5 FPGA as the target device. The simulation results are matched favourably with that of the output obtained by MATLAB R2010a with MSE of1.3861x10-4.