High performance frame selection algorithm for gray-level frames within the framework of multi-frame super-resolution

Abstract

Multi-frame image super-resolution represents an efficacious albeit expensive and resource-intensive technique for image reconstruction, necessitating substantial memory allocation for data storage. To mitigate the computational burden inherent in multi-frame image super-resolution algorithms, a strategic approach involves curtailing the processing load by disregarding redundant frames. In this study, we introduce a novel frame selection algorithm tailored to identify an optimal minimum number of frames. This approach ensures the fidelity of the reconstructed high-resolution (HR) image while significantly alleviating the procedural complexity and memory demands of the super-resolution process. The frame selection methodology we propose is founded upon multi-channel sampling, reference frame selection, and the maximization of the lower bound on the signal-to-noise ratio. Specifically, our approach is operationalized through two optimization algorithms based on priority search. The initial algorithm identifies cases with maximum non-empty channels by exploring the predefined domain encompassing all feasible desired positions. In the subsequent algorithm, the process entails identifying, for any channel within any discovered case, a frame associated with the minimum translation function model noise. Subsequently, the total noise of each case is computed. We ascertain the optimal case along with a collection of frames that correspond to the minimum total noise. Experimental findings highlight the efficacy of our proposed method in mitigating super-resolution complexity while achieving high-fidelity HR images that closely match or surpass those generated from complete frame sets. Comparative analysis against established super-resolution (SR) algorithms demonstrates the remarkable speed and minimal computational overhead of our proposed approach, rendering it exceptionally efficient with negligible runtime. © 2025 Elsevier Inc.