A neural approach to optical image reconstruction

The paper presents, with experimental results, a method of applying Fourier optical signal processing as a pre-processor to a digital signal system. The input illumination is a coherent light source from a HeNe laser and computation of the Fourier transform (FT) is carried out via a FT lens. An image is placed in the front focal plane of the FT lens and the Fourier transform appears at the rear focal plane. A low cost charged coupled devices (CCD) camera is employed to capture the optical Fourier signal. Due to avalanche effects and the easily saturated characteristics inherent in all CCD cameras, a very noisy and saturated power spectral density is captured. In order to retrieve the original image back using the input/output approach proposed by Fienup [1978], important criteria like the initial guess object and the imposed Fourier object constraints plays a crucial role in the reconstruction process. The paper also proposes a low-cost and an efficient way of how neural networks can be used as a productive tool in the process of solving the phase-retrieval problem of reconstructing a general object from the modulus of its Fourier-transformed optical image. Problems encountered during the construction phase were studied and solutions were provided both at the optical processing and digital system ends. This helps to understand the accuracy of the transformation, the practical behaviour and characteristics of the optical lens system