Multi-stage progressive single image joint motion deblurring and super-resolution

• Main Authors: Shoyan, Mekhak, Shoyan, Misak, Navasardyan, Shant, Hakobyan, Robert
• Format: Conference Proceeding
• Language: English
• Subjects: Artificial neural networks; Blurring; Image resolution; Image restoration; Neural networks; Signal to noise ratio; Spatial data
• ISSN: 0094-243X; 1551-7616
• DOI: 10.1063/5.0136155

In the past decade, deep convolutional neural networks achieved great success in the task of single image super-resolution providing realistic texture and structure details. However, sometimes low-resolution images along with a lack of spatial information are degraded by complicated blur effects. Especially when images are taken in a moving environment or with camera shakes, one may end up with results suffering from motion blur. In this work, the joint problem of single image motion deblurring and super-resolution is addressed. We propose a single branch fully convolutional neural network to restore high-resolution sharp images from given motion blurry low-resolution images without estimating or making any assumptions on the blur uniformity, its spatially varying filter, and noise. Our end-to-end solution reuses the features extracted from the motion deblurring network in the super-resolution module resulting in an efficient model with high performance. Experiments show that our method outperforms the existing state-of-the-art solutions qualitatively and quantitatively by reaching 28.265 peak signal-to-noise ratio. Our implementation code is available at https://github.com/Mekhak/motion-deblur-sr.