Comparative Analysis of Quantization Optimizations for Single-Image Super-Resolution Models

This study describes how to increase inference performance without critical image quality loss by using post-training quantization methods for lightweight single-image super-resolution (SISR) models. To observe this improvement, we compare the performance of the quantized models with the original models. Benchmark tests show that 8-bit quantization accelerates the original SISR models by 1.3x to 2.7x on an energy efficient GPU computing unit for various datasets. In our experiments, we compare the peak signal-to-noise ratio (PSNR) and inference times of commonly used activation functions using an interoperability framework. According to the test results, the quantized model with ReLU accelerates the inference compared to the quantized model that originally contained Tanh.