Radon transform computations using DSP chips: an evaluation and comparison

A description is given and a comparison is made of different interpolation schemes for the computation of the Radon transform and backprojections. The nearest neighbor method is fastest, but the resulting reconstructed image is not very good. For D=1 the best method to use is linear interpolation because its performance is virtually identical to the line-length method at a reduced computational complexity. If D>1, the best reconstruction is obtained using the line-length method. The authors compared implementations of these procedures on two DSP chips. Although the DSP 16 clock cycle time is only 55 ns, which is 3.6 times faster than the TMS32020 200-ns clock cycle time, it performs backprojection computations only about 1.8 times faster than the slower TMS32020 chip. The 60-ns cycle-time TMS320C30 is capable of performing the Radon and inverse transforms faster than the 55-ns DSP16, due to the differences in instruction sets and architectures of the chips