Generative imaging and image processing via generative encoder

<p style='text-indent:20px;'>This paper introduces a novel generative encoder (GE) framework for imaging and image processing tasks like reconstruction, compression, denoising, inpainting, deblurring, super-resolution. GE unifies the capacity of GANs stability AEs in an optimization instead stacking into single network or combining their loss functions as existing literature. provides approach to visualizing relationships between latent spaces data space. The is made up pre-training phase solving phase. In former, GAN with generator <inline-formula><tex-math id="M1">\begin{document}$ G $\end{document}</tex-math></inline-formula> capturing distribution given set, AE id="M2">\begin{document}$ E that compresses images following estimated by id="M3">\begin{document}$ are trained separately, resulting two representations data, denoted encoding space respectively. phase, noisy id="M4">\begin{document}$ x = \mathcal{P}(x^*) $\end{document}</tex-math></inline-formula>, where id="M5">\begin{document}$ x^* target unknown image, id="M6">\begin{document}$ \mathcal{P} operator adding addictive, multiplicative, convolutional noise, equivalently such id="M7">\begin{document}$ compressed domain, i.e., id="M8">\begin{document}$ m E(x) unified via problem</p><p style='text-indent:20px;'><disp-formula> <label/> <tex-math id="FE1"> \begin{document}$ z^* \underset{z}{\mathrm{argmin}} \|E(G(z))-m\|_2^2+\lambda\|z\|_2^2 $\end{document} </tex-math></disp-formula></p><p style='text-indent:20px;'>and id="M9">\begin{document}$ recovered way id="M10">\begin{document}$ \hat{x}: G(z^*)\approx id="M11">\begin{document}$ \lambda>0 hyperparameter. unification allows improved performance against corresponding networks while interesting properties each space.</p>