A real-time implementation of Richardson-Lucy deconvolution

Deconvolution is an important technique in image processing that may be used to recover images that have been subjected to a blurring process, usually caused by atmospheric effects or limitations of the image capturing equipment. Noise in the image data means that the problem is ill-posed, and thus mathematically complex statistical estimation techniques must be employed. This complexity, and the high throughput levels required for video data, renders a real-time software implementation unfeasible, however the parallelism of FPGA devices makes them an ideal medium. In this paper an FPGA implementation of an accelerated Richardson-Lucy deconvolution algorithm will be presented. The design uses multistage separable filters as a hardware efficient means of implementing the several large 2D convolutions that are required. The results show that real-time full scene deconvolution is viable with today's FPGA technology.