Efficient Deep Learning based Super-Resolution via Dynamic Quantization

Abstract

Efficient Deep Learning based Super-Resolution via Dynamic Quantization Jaemin Park Dept. of Intelligence and Convergence The Graduate School Hanyang University Although the significant advances in image super-resolution (SR) with convolutional neural networks have led to continuous performance improvement, the state-of-the-art (SOTA) SR networks face the challenges conducting in resource constrained environments due to its complexity. To address these challenges, one of well-known neural network optimizations, quantization, have been widely employed. However, when quantizing the SR networks, their characteristics should be taken into account. To tackle this issue, this work proposes the efficient deep learning-based SR via dynamic quantization. This approach takes into account the frequency of the input image and layers to determine the adequate bit-widths for quantizing each layer. Frequency is the fundamental element in images that has been shown to enhance the SR performance in several recent works [1,2]. Moreover, in order to differentiate between the quantization features of different bit- width and achieve the SR performance similar to that of original full precision network, this work has incorporated the multiple teacher knowledge distillation (KD) with contrast loss into the framework. This quantization approach has been implemented across various widely used SR networks such as EDSR [3] and IMDN [4]. Experiments show that suggested framework achieves a superior performance-to-compression ratio compared to SOTA frameworks for CNN-based SR model quantization.