A novel energy-efficient edge computing paradigm is proposed for real-time\ndeep learning-based image upsampling applications. State-of-the-art deep\nlearning solutions for image upsampling are currently trained using either\nresize or sub-pixel convolution to learn kernels that generate high fidelity\nimages with minimal artifacts. However, performing inference with these learned\nconvolution kernels requires memory-intensive feature map transformations that\ndominate time and energy costs in real-time applications. To alleviate this\npressure on memory bandwidth, we confine the use of resize or sub-pixel\nconvolution to training in the cloud by transforming learned convolution\nkernels to deconvolution kernels before deploying them for inference as a\nfunctionally equivalent deconvolution. These kernel transformations, intended\nas a one-time cost when shifting from training to inference, enable a systems\ndesigner to use each algorithm in their optimal context by preserving the image\nfidelity learned when training in the cloud while minimizing data transfer\npenalties during inference at the edge. We also explore existing variants of\ndeconvolution inference algorithms and introduce a novel variant for\nconsideration. We analyze and compare the inference properties of\nconvolution-based upsampling algorithms using a quantitative model of incurred\ntime and energy costs and show that using deconvolution for inference at the\nedge improves both system latency and energy efficiency when compared to their\nsub-pixel or resize convolution counterparts.\n