Lightweight Super-Resolution for Real-World Burst Images Captured by Handheld Camera Sensors Based on Partial Differential Equations

Due to the limited specifications of smartphone cameras, such as small sensors that restrict spatial resolution, the primary goal of the image processing unit (IPU) within smartphones is to enhance image quality and produce high-resolution outputs. In addition, modern smartphone cameras feature a default burst photography mode that captures consecutive photos of the same scene in quick succession. Since different burst images contain varying undersampled and shifted information, utilizing and combining this information from multiple burst images to produce a single high-resolution image has become a natural approach. With the success of deep learning in image processing, numerous deep learning-based burst imaging methods have recently been proposed. However, learning-based burst image restoration still faces challenges, such as the difficulty of handling varied conditions like low lighting and complex hand movements, the need for extensive training data, and the requirement of significant memory resources. Consequently, their results on real-world burst images often exhibit color artifacts like moiré patterns and cannot reconstruct small structures well. In this paper, we propose a non-deep learning-based method for burst-image super-resolution that employs global three-dimensional alignment along with a partial differential equation framework. The proposed method is well-suited to real-world burst images, effectively reduces color artifacts, even those introduced by state-of-the-art deep learning methods, and achieves fine structural reconstruction that is challenging for these deep learning-based approaches.