A generalized modeling of ill-posed inverse reconstruction of images using a novel data-driven framework

By definition, an instance of image reconstruction often combines denoising, deblurring and enhancing objectively and appears as a fundamental process in visual processing systems which is mathematically an ill-posed inverse optimization problem. It suffers additional complexities because of a non-stationary nature of the image, restricting available methods to produce inconsistent restoration on a variety of image and degradation types. It requires filling this gap by a generic framework of image restoration. Therefore, in this paper, we modeled a novel data-driven framework of generic image reconstruction as a testing benchmark to the emerging application of deep learning. The proposed framework comprises of data engineering, image filtering and a regression neural network for better restoration of many degraded images. We evaluated the effectiveness of the framework on many images degraded by a Gaussian, out-of-focus, motion or airy-pattern blur and random additive white Gaussian noise. The performance appeared better in comparison to a benchmark and state-of-the-art generic image restoration results on several indicators. Furthermore, the framework performed better in comparison to a recent approach to image reconstruction which uses a deep convolutional neural network.