next up previous print clean
Next: Equivalence of filter- and Up: Processing local-stationary data Previous: Interpolation

Implementation and Interface

The programmer interface of the Patch classes successfully encapsulates and hides the somewhat complex implementation and invocation details of the underlying patch enumeration, windowing, weighting, and merging. The program NStatVarPush[*] shows the instantiation and invocation of the patch method. In general, The programmer chooses among two patch operators: PatchPush or PatchPull. PatchPush defines the patches as subsets of the input, while PatchPull defines them as subsets of the output. The initialization of the patch operators requires three arguments: the patch size and overlap, the input or output space, and an operator factory object. The patch operators implement Jest's operator interface Schwab and Schroeder (1997) and,consequently, the standard image() method accepts the input quilt and returns the processed output quilt.


package nstat.var;
import  jam.operator.*; 
import  rsf.vector.*;
import  rsf.operator.*; 
import nstat.patch.*;

public class NStatVarPush { public static void main(String[] args) { ArgumentKeeper argKeeper = new ArgumentKeeper(args); // read args Rsf iput = RsfFactory.newRsf ( ); // read input quilt float olap = argKeeper.getFloat ("overlap" , "0.5"); int[] psze = argKeeper.getIntArray("patchsize", "321");

OperatorFactory opFac = new StatVarFactory(); PatchPushOp pchOp = new PatchPushOp((RsfSpace) iput.getSpace(), olap, psze, opFac); Rsf oput = (Rsf) pchOp.image(iput); // image quilt oput.write("pushOput.H"); iput.write("pushIput.H"); // write output quilt } }

The patch operators require an operator factory to generate the stationary operators needed to process each stationary patch. Such a factory is usually easily programmed, because most of its methods are simply wrappers for the instantiation methods of the stationary operator. Why are such wrappers necessary? A factory standardizes instantiation methods, which cannot be standardized directly by inheritance from an abstract interface (). Additionally, the operator factory supplies methods that return the domain given a range and vice versa. Table 1 shows the basic methods of the operator factory interface.

Table 1: Some Operator Factory methods (syntax, effect).
xspc = fac.createDomain(yspc) creates domain space for given range space  
yspc = fac.createRange( xspc) creates range space for given domain space  
oper = fac.getOp(xspc,yspc) creates operator given domain and range  

The programmer-friendly patch operators encapsulate the tedious and error-prone computations that split, process, and merge nonstationary data sets[*]. Internally, the patch operators invoke a class called PatchEnumeration that organizes the patch processing. The enumeration computes the actual patch location. The patch size is kept constant, but the overlap may slightly vary in order to fit the given patch size into the given input quilt size. The patch location is defined by the patch size and a patch anchor, the lowest-index vertex of the patch on the quilt space. The PatchEnumeration object directs a Window operator ${\bf P}_i$ to extract the various patches. For efficiency, the Window operator can change its domain and range without being instantiated. The operator handles minor rounding errors in the spaces' offset and increment intelligently, so that the patch's space remains a subspace of the quilt space. For each extracted patch, the patch operator requests a corresponding operator ${\bf F}_i$ from the operator factory. After processing the patch, the patch operator creates an interpolation operator ${\bf Q}_i$. The interpolation operator weights the output patch and adds it to the output quilt ${\bf m}$. The same interpolation operator images an identity vector and adds it to the weight quilt $\hat{\bf w}_m$. After the last patch is processed, the patch operator divides the output quilt's elements by the corresponding elements of the weight quilt. The encapsulation of the patch operator greatly owes its implementation to Jest's novel and careful formulation of domains, ranges, and vector spaces.

next up previous print clean
Next: Equivalence of filter- and Up: Processing local-stationary data Previous: Interpolation
Stanford Exploration Project