Reinforced Hybrid Wiener Deconvolutional - Convolutional Autoencoders Based Image Deblurring

This paper is addressing the issue of image deblurring by employing a novel reinforced hybrid Wiener deconvolutional-convolutional neural network (HWDCNN) in the context of images affected by Gaussian blur. The proposed method is set to explore the capabilities of a custom NN layer that performs the classical Wiener deconvolution operator in a trainable manner, hence, leading to a blind deblurring method where the weights represent the point spread functions (PSF) and signal-to-noise ratios (SNR) of the deconvolution layer which are optimized in a traditional deep learning paradigm. Additionally, the custom layer is set as a first layer on two of the branches of the model, followed by layers of transposed convolution to explore possible synergies. The results indicate that this method outperforms known state-of-the-art methods for the presented dataset setup in terms of both structural similarity index measure (SSIM) and peak SNR (PSNR) metrics.