{"id":351,"date":"2013-09-14T12:24:00","date_gmt":"2013-09-14T12:24:00","guid":{"rendered":"http:\/\/ahay.org\/blog\/?p=351"},"modified":"2015-08-30T20:57:56","modified_gmt":"2015-08-30T20:57:56","slug":"program-of-the-month-sfpatch","status":"publish","type":"post","link":"https:\/\/ahay.org\/blog\/2013\/09\/14\/program-of-the-month-sfpatch\/","title":{"rendered":"Program of the month: sfpatch"},"content":{"rendered":"

sfpatch<\/a> breaks the input data into local windows or “patches”, possibly with overlap. <\/p>\n

The patching technique is explained by Jon Claerbout in Nonstationarity: patching<\/a> chapter from Image Estimation by Example<\/em>. <\/p>\n

\"\" <\/p>\n

Suppose you have a 1-D signal with 10 samples: <\/p>\n

<\/i>
<\/i><\/a><\/div><\/div>
bash$ sfmath n1=10 output=x1 > data.rsf\n<\/code><\/pre><\/div>\n
You can divide it, for example, into two patches with 5 samples each: <\/p>\n
<\/i> 
<\/i><\/a><\/div><\/div>
bash$ < data.rsf sfpatch p=2 w=5 > patch.rsf\nbash$ sfget n1 n2 < patch.rsf \nn1=5\nn2=2\n<\/code><\/pre><\/div>\n
or into 5 overlapping patches with 3 samples each: <\/p>\n
<\/i> 
<\/i><\/a><\/div><\/div>
bash$ < data.rsf sfpatch p=5 w=3 > patch.rsf\nbash$ sfget n1 n2 < patch.rsf \nn1=3\nn2=5\n<\/code><\/pre><\/div>\n
If you specify only the patch size (w=<\/strong> parameter), sfpatch<\/strong> tries to select the number of patches to achieve a good overlap: <\/p>\n
<\/i> 
<\/i><\/a><\/div><\/div>
bash$ < data.rsf sfpatch w=3 > patch.rsf\nbash$ sfget n1 n2 < patch.rsf \nn1=3\nn2=6\n<\/code><\/pre><\/div>\n
If you put overlapped patches back together, their amplitudes add in the overlapped regions: <\/p>\n
<\/i> 
<\/i><\/a><\/div><\/div>
bash$ < data.rsf sfdd type=int | sfdisfil \n   0:    0    1    2    3    4    5    6    7    8    9\nbash$< patch.rsf sfpatch inv=y | sfdd type=int | sfdisfil\n   0:    0    2    6    6    8   10   12   14    8    9\n<\/code><\/pre><\/div>\n
unless you use the weight option: <\/p>\n
<\/i> 
<\/i><\/a><\/div><\/div>
bash$ < patch.rsf sfpatch inv=y weight=y | sfdd type=int | sfdisfil\n   0:    0    1    2    3    4    5    6    7    8    9\n<\/code><\/pre><\/div>\n
bash$ sfpatch<\/strong> easily handles multidimensional data: w=<\/strong> and p=<\/strong> become lists of dimensions, and the number of dimensions in the output effectively doubles. sfpatch<\/strong> is useful in two applications: <\/p>\n
    \n
  1. crude handling of non-stationarity when processing non-stationary signals with locally stationary filters,<\/li>\n
  2. making non-parallel tasks data-parallel.<\/li>\n<\/ol>\n
    A simple example from rsf\/rsf\/mona<\/a> illustrates the second use. <\/p>\n
    \"\" \"\"<\/p>\n
    Anisotropic diffusion, implemented by sfimpl2<\/a>, is an effective but relatively slow process. By breaking the input 512×512 image into nine 200×200 overlapping patches and by processing them in parallel on a multi-core computer, we can achieve a significant speed-up without rewriting the original program. <\/p>\n
    10 previous programs of the month<\/h3>\n