A Functional Approach to Border Handling in Image Processing

Domain Specific Languages for image processing offer simplified programming and efficient execution on a variety of platforms. These languages are particularly suitable for implementing local image processing operators that produce each output pixel by processing a corresponding window of input pixels. Such operators are easily parallelised and vectorised for efficiency. However, border handling is required to define the results for pixels close to the image border where the input window partially falls outside the bounds of the input image. We propose a declarative approach to border handling in a functional image processing Domain Specific Language called Halide*. The compiler uses code analysis to infer the bounds of output images and to optimise border padding implementations. Experimental results demonstrate good execution efficiency for large images.