454 technology produces a number of errors in the reads, mostly (but not only) related to homopolymeric runs. It requires some degree of quality filtering, that is removing reads that contain false information. It's often based on quite simple measures of number of consecutive low quality bases and length. Are there any other approaches to quality filtering than the ones implemented in Pyro/AmpliconNoise packages?
If you have a reference or high coverage then using the 454 toolchain for mapping or assembly should handle this. The specific error modality you refer to, homopolymer runs, does not require the removal of the reads but careful calling of certain bases (those in homopolymer runs). Recent versions of the Newbler software output a histogram of signal strengths for the homopolymer runs to allow you to see the distribution of signal at these sites.