Minimum entropy-neural network approach to turbulent-image reconstruction

We investigate a neural net-based algorithm for enhanced imaging through atmospheric turbulence. The concept is based on a standard model of optical turbulence, according to which a short-exposure point-spread function is a random superposition of speckles. This leads to a new method of image processing called the Fourier division approach. The latter requires the taking of two short-exposure images in rapid succession, which are picked up by an image-plane array, divided in Fourier space, and then processed by a minimum entropy-neural net approach. The main task of the processing is to estimate the two short-exposure point-spread functions that characterize the two images. Given these estimates, the two images may now be inverse filtered to produce two sharp object-scene estimates. These have most of the turbulence degradation removed, and are averaged to produce a single output image. The approach shows promise, in computer simulations, of removing nearly all of the turbulence degradation very quickly (currently tens of seconds). A further benefit arises from knowledge of the twoshort-exposure point-spread functions. These should permit identification of the state of turbulence along the imaging line of sight and, in particular, the presence of wind shear.