Abstract:
In recent years, deep learning-based single image super-resolution (SISR) has attracted
considerable attention and achieved significant success on advanced GPUs. Most state-
of-the-art methods require a large number of parameters, memory, and computational
resources, often resulting in inferior inference times on mobile devices.
In this thesis, we introduce a plain convolution network augmented with a nearest-
neighbor convolution module and 8-bit quantization to achieve real-time SISR on NPUs.
Furthermore, we evaluate the efficiency of our network architecture by comparing ex-
periments on mobile devices to select the tensor operations to implement. The model
comprises only 52 K parameters, achieves 4× upscaling in 0.065 s on a Snapdragon
865 CPU smartphone, and by comparing to other SR methods, we found that our model
can achieve high fidelity super resolution results while using fewer inference times.
