Regularization and Applications of a Network Structure Deep Image Prior

Finding a robust image prior is one of the fundamental challenges in image recovery problems. Many priors are based on the statistics of the noise source or assumed features (e.g. sparse gradients) of the image. More recently, priors based on convolutional neural networks have gained increased attention, due to the availability of training data and flexibility of a neural network-based prior. Here, we present results for an entirely novel neuralnetwork based image prior that was introduced in [18], known as a network structure deep image prior. We test the effect of regularization on convergence of the network structure prior, and apply this prior to problems in deconvolution, denoising, and single-pixel camera image recovery. Our results show that regularization does not improve the convergence properties of the network. The performance for improving deconvolution results and single-pixel camera image recovery are also poor. However, the results for denoising are comparable to baseline methods, achieving a strong PSNR for several test images. We believe with further work, this method can be readily applied to similar computational imaging problems such as inpainting and demosaicking.