@Hui_Yang, can you provide explicit sequence examples where you think the reads are simply being stitched together end-to-end? This should not be the case, especially as there must be a minimum overlap of 10 bases for a successful merge in vsearch (default), unless you are changing this value. While DADA2 requires a 12 base overlap, (currently cannot be altered via the dada2 plugin at the command-line interface).
What would be most helpful is the output from the --o-denoising-stats of DADA2 or --o-stats from deblur. These outputs will inform you of where most of your data is being lost. That is, mergning, denoising etc...
-Mike