Guided super-resolution

The image formation process by a remote sensor is modeled, highlighting restrictions on spatial resolution and image formation artifacts. The super-resolution problem is introduced, describing the aim of overcoming the resolution and image quality limits imposed by the sensor optics and sampling. Problems with model inversion and conventional image restoration techniques are discussed.; The focus of this thesis is a Super-Resolution technique in which image geometry is used to guide the resampling. The dynamics considered are those of curvature flow, in which image restoration—including Fourier mode synthesis, noise-supression, and anti-aliasing—is achieved my minimizing the mean curvature of an image. Following a discussion of the curvature flow switch introduced by Malladi and Sethian, we cast their work into the neural regime, identifying a similarity to the work of Tatem et al.; The performance of the switch is compared to the performance of polynomial interpolation. Testing revealed excellent performance: the super-resolved images are sharp with high signal to noise ratio relative to polynomial-interpolated images. The peak signal-to-noise ratio is typically greater than that of polynomial-interpolated images by ten to thirty percent. The advantage in many spatial frequency bands—often the highest frequency bands—reaches several hundred percent.