Next: , Previous: , Up: Internal Routines   [Contents][Index]

#### 15.5.535 segment

```segment(data [, sign=sign, diagonal=diagonal, /degree])```

Curvature-based data segmentation in multiple dimensions. `segment(data)` returns a `1` for each element of `n`-dimensional `data` where the data in an n-dimensional hypercube with sides of 3 elements centered on the element under investigation shows curvature of the appropriate sign in all directions through that central element, and a `0` otherwise. For each direction, the curvature is equal to the sum of the values of two elements adjacent on opposite sides to the central element minus twice the central value. For example, in two dimensions we can regard the data as an altitude map, and then this function returns a `1` in all places where the central value exceeds the average of the two adjacent values on opposite sides of the central value in the n-s direction, the e-w direction, the ne-sw direction, and the nw-se direction. The number of directions to consider per data element in an `n`-dimensional array `data` is equal to ```m = (3^n - 1)/2```.

If `diagonal` is specified, then it must have one element per data dimension and is taken as instructions on which nearest-neighbor connections to recognize. A `2` means that nearest neighbors with their `bitmap` values equal to `1` are assigned to the same area if they share any face or vertex in the corresponding dimension. A `1` means that such neighbors are assigned to the same area only if they share a face in the corresponding dimension. a `0` means that connections in the corresponding dimension are ignored. If `diagonal` is not specified, then code `2` is assumed for all dimensions.

For example, if `diagonal` is equal to `[2, 2, 0]`, then no connections in the third dimension are recognized, and the data is treated as a set of unconnected two-dimensional images (but no area in any of these images shares its area number with any area in any of the other images). Within each of the images, connections in all four directions (n-s, e-w, ne-sw, nw-se) are recognized in this case. If `diagonal` is equal to `[1, 1, 0]`, then only directions in the orthogonal directions (n-s, e-w) are recognized. Likewise for `[2, 1, 0]`, because even though the first dimension allowes vertices, the second one only accepts faces.

If `/degree` is specified, then the number of directions per data element that exhibit the appropriate curvature is returned instead. This number always lies between 0 and the total number `m` of directions given above. `segment(data, /degree) eq m` is equivalent to `segment(data)`, but the latter form is usually considerably faster because for each element it stops checking as soon as a direction with inappropriate curvature is found. For noise data in arbitrary numbers of dimensions, on average at most 2 curvature calculations per data element are performed.

If `sign` (when converted to type `long`) is non-zero, then negative curvature (i.e., hill-like objects) are sought. If integer `sign` is negative, then positive curvature (i.e., valley-like objects) are sought. By default, a value of `+1` is assumed for `sign`.

Zeros are returned on the 1-px-wide edges of the result, because not enough data is available there to determine the curvature in all directions.

If `data` is a 2-dimensional array and neither `diagonal` nor `/degree` are specified, then a special-purpose version of the routine is used that is optimized for processing speed. In all other cases a more general but slower version is executed.

If `data` is of types `long`, `float`, or `double`, then overflow problems may occur if any of its values have a magnitude exceeding `#max_long/2`, `#max_float/2`, or `#max_double/2`, respectively.

Next: , Previous: , Up: Internal Routines   [Contents][Index]