Len Hamey
Macquarie University

Domain-Specific Characteristics of a Machine Vision Library

This talk will explore the domain-specific language features required to
implement the image processing and low-level computer vision functions in a
commercial machine vision library, and the implications for an
image-processing/low-level computer vision DSL such as Apply.  Of
approximately 1200 functions in the library, 260 were image related, the
rest being concerned with other data structures, system interface, etc.
Examining these functions, we identified 6 useful language features not
currently implemented in Apply: reducers, variable kernel size, convolution
mask calculation, image warping, raster cursive computation and multi-pass
operations.  Based upon our analysis of the library, we quantified the
contribution of each feature in terms of its benefits for re-implementing
the library's functions.  The talk will discuss the proposed language
features and consider their implications for efficient parallel compilation
of Apply modules.  A key contribution of this talk is the identification and
prioritisation of language capabilities required for efficient programming
of image processing and low-level computer vision functions.

This talk is based on work to be presented at DICTA 2007.