Blind deconvolution network of encoded images in wavefront coding systems

An optical-digital imaging system that uses some Jacobi-Fourier profile functions for encoding and a convolutional neural network (<i>CNN</i>) for digital decoding is proposed. A blind deconvolution network is trained on generated encoded image datasets with defocus aberration added to the in-focus point spread function (<i>PSF</i>) corresponding to each phase mask (<i>PM</i>). Image encoding is done using a <i>PSF</i> that belongs to the classic cubic <i>PM</i> and the family of <i>PM</i>s based on the Jacobi-Fourier polynomials (<i>JFP</i>), with radial order p= 7, 9, and 10. Our simulated and experimental results show that the proposed trained model can recover high frequencies in images acquired under different defocus values without knowing the exact optical transfer function (<i>OTF</i>) of the optical system. Unlike other approaches, no preprocessing steps, such as noise removal, radiometric normalization, or spatial enhancement, are performed on the experimental data. Also, our blind deconvolution network can decode optically encoded images, although none belong to the trained dataset. It is helpful in practical applications, where minor differences between the real <i>OTF</i> of the imaging system and the <i>OTF</i> used in a deconvolution algorithm significantly impact the final image quality. As is shown, our basic <i>CNN</i> architecture can recover a final image with minimal artifacts and higher contrast at different defocus values without considerable computational resources.