DWSR: an architecture optimization framework for adaptive super-resolution neural networks based on meta-heuristics

Abstract

AbstractDespite recent advancements in super-resolution neural network optimization, a fundamental challenge remains unresolved: as the number of parameters is reduced, the network’s performance significantly deteriorates. This paper presents a novel framework called the Depthwise Separable Convolution Super-Resolution Neural Network Framework (DWSR) for optimizing super-resolution neural network architectures. The depthwise separable convolutions are introduced to reduce the number of parameters and minimize the impact on the performance of the super-resolution neural network. The proposed framework uses the RUNge Kutta optimizer (RUN) variant (MoBRUN) as the search method. MoBRUN is a multi-objective binary version of RUN, which balances multiple objectives when optimizing the neural network architecture. Experimental results on publicly available datasets indicate that the DWSR framework can reduce the number of parameters of the Residual Dense Network (RDN) model by 22.17% while suffering only a minor decrease of 0.018 in Peak Signal-to-Noise Ratio (PSNR), the framework can reduce the number of parameters of the Enhanced SRGAN (ESRGAN) model by 31.45% while losing only 0.08 PSNR. Additionally, the framework can reduce the number of parameters of the HAT model by 5.38% while losing only 0.02 PSNR.