Noise-insensitive no-reference image blur estimation by convolutional neural networks

A few image quality metrics for blur assessment have been presented in the last years. However, most of those metrics do not take image noise into account. Yet, image noise is an unavoidable part of the image forming process with digital cameras. Some thermal imagers show larger sensor noise and inhomogeneity compared to cameras operating in the visible range. Further, natural imagery might contain a combination of several degradations. Assessment of degraded images by observer trials is expensive and time consuming. A single robust quality metric might be derived by metrics highly responsive to single degradations and insensitive to others. Hence separate assessment of image blur and noise seems to be reasonable. In this paper we present a deep learning approach for noise-insensitive blur predictions by using Convolutional Neural Networks (CNN) on image patches. In contrast to current blur metrics the model output is highly correlated to blur distortion over a wide range of image noise. The model is trained on images of ImageNet database impaired by Gaussian blur and noise and tested on artificial and natural image data. Local blur estimation based on patches is especially useful for estimation of non-uniform blur due to motion and atmospheric turbulence.