Neural network non-uniformity correction by means of frames motion estimation and its hardware implementation

The non-uniformity in the infrared focal plane array has limited the quality of infrared imaging system. Scene-based neural networks (SBNNT) non-uniformity correction (NUC) techniques correct the non-uniformity by using an image sequence and relying on motion between frames. An improved SBNNT non-uniformity correction of IRFPA is proposed to eliminate the ghost artifact. Linear interpolation projection-based shift estimation (LIPSE) algorithm is used to select frames for learning and the offset matrix of SBNNT was calculated or updated depending on the reference flame numbers. The improved algorithm has run on a small low power consumption DSP hardware platform with TMS320DM643 as the kernel processor. The result shows that the non-uniformity correction can be realized in a simple way with satisfactory.