Averaging approaches for highly accurate image-based edge localization

Abstract

We introduce an optical and a digital averaging technique that considerably improves edge localization performance. Especially for high quality images, the optical method achieves measurement uncertainties down to levels of millipixels. The approach uses an optical replication scheme based on a computer-generated hologram to reduce noise and discretization errors. The second method is based on a neural network denoising architecture and is especially suited for high levels of photon noise. Edge localization can be improved by up to 60% while preserving high lateral and temporal resolution. The methods are first tested using high quality images obtained by a scientific CMOS sensor imaging a razor blade mounted on a mechanical stage. Then, the laboratory results are tested for larger distances to validate the methods for building deformation measurements.