Mobile Edge Computing (MEC) has been proposed to enhance the performance of Internet of Things (IoTs) devices by offloading computation-intensive tasks to nearby edge clouds, while Non-Orthogonal Multiple Access (NOMA) enables multiple IoTs devices to share subcarriers with varying power levels, making it ideal for computation offloading. Despite the potential benefits, integrating NOMA with MEC presents complex challenges, including resource allocation, decision optimization, and balancing energy efficiency with completion time. In this paper, we address the computation offloading and resource allocation problem in NOMA-MEC enabled IoT networks, aiming to minimize completion time and maximize energy efficiency while meeting processing latency requirements. Our model supports partial computation offloading, allowing devices to partition tasks for both local execution and offloading to the edge clouds. To this end, we first introduce two processes, i.e., infeasible tasks elimination and admission control, to improve algorithm efficiency. Then, we propose an iterative algorithm, comprising two low-complexity sub-algorithms, to address various optimization aspects, including CPU frequency allocation, offloading decisions, time allocation, transmit power control, and network resource allocation. Extensive simulations validate that our approach outperforms existing methods in terms of completion time, total saved energy, and offloading ratio.

Video frame interpolation (VFI) seeks to enhance visual quality by synthesizing intermediate frames, effectively increasing the frame rate. While existing supervised VFI methods have shown remarkable success, they are heavily reliant on large, labeled datasets of high frame-rate videos, posing significant challenges in terms of acquisition cost and scalability. To address this limitation, we present VFI-E (Video Frame Interpolation by Extrapolation), a novel unsupervised approach that leverages frame extrapolation as a self-supervisory signal for high-quality VFI. Our central hypothesis is that the fidelity of an interpolated frame can be effectively gauged by its ability to facilitate accurate frame extrapolation. VFI-E operates by first synthesizing an intermediate frame using a flow-based interpolation network. This interpolated frame, along with one of the input frames, is then fed into an extrapolation network tasked with reconstructing the other input frame. Crucially, the reconstruction loss between the original and extrapolated input frames serves as the driving force for optimizing the interpolation network, removing the need for ground-truth intermediate frames. Extensive experiments on benchmark datasets, including UCF101, Vimeo-90k, and Adobe240-fps, demonstrate that our unsupervised VFI-E achieves performance comparable to state-of-the-art supervised methods. Specifically, VFI-E attains competitive PSNR and SSIM values on single-frame interpolation tasks and exhibits strong performance on multi-frame interpolation, showcasing its effectiveness and generalization capabilities.