Adaptive Infrared Non-Uniformity Correction

Real-time adaptive non-uniformity correction by a neural network algorithm was implemented on the 12- bit digital image from a Boeing SE-U20 uncooled 320x240 microbolometer camera. Nonlinearities in an infrared sensor require either periodic recalibration of a one or two point correction algorithm as the scene and environment change or require an adaptive continuous correction. The adaptive neural network correction is performed in real-time with an off-the-shelf processor board inserted in an IBM PC compatible machine. The real time implementation allows long term stability and performance issues of the algorithm to be addressed. Evaluation of the adaptive algorithm shows that the spatial noise in the corrected image depends strongly on the estimate of the desired image used in the adaptive algorithm. The desired value is calculated by means of neighborhood functions such as median or convolution with kernels such as the sinc function. We have determined that the adaptive algorithm works better when the time sample between images of a moving scene is large; that is, when the images are relatively uncorrelated. This effect must be balanced by the need to have the algorithm converge in a finite time. The net effect of this balance is that the hardware signal processing requirements are reduced considerably since the algorithmic calculations need not be done on every frame.