The image processing functions for single images fall into six main classes: intensity transformations, local neighborhood filtering (convolutions), image warping, Fourier filtering, statistical filtering, and morphological transformations. The TMC 2301 supports local kernel (up to 16×16) neighborhood weighted convolutions with or without coordinate transformation. For every output pixel (using a pixel-filling approach) a walk for N×N pixels surrounding the corresponding input position is generated and used to calculate the convolution with a user-specified table of coefficients or weights.
This processor, with an address-calculation, I/O, and convolution module, is therefore suited for both conventional image processing board functions and the more costly image warpings (scale, rotate, and translate). For image warpings, output intensity values can be obtained using nearest neighbor, bilinear interpolation, or cubic convolution.
The paper describes the use of a TMC 2301 within gray-level or true-color image processing systems and gives examples of how to program the TMC 2301 for a number of image warping and convolution filters down to the lowest detail. As such, it is appropriate, well-written, and well-illustrated. Regrettably, it lacks references to or speed and price comparisons with related architectures offering the same functionality.