To tackle the issue of parameter redundancy and the challenge of learning comprehensive global contextual information in current deep learning-based super-resolution methods,as well as the limited capacity to reconstruct high-frequency image features,a super-resolution network(RPSRnet)based on receptive field optimization and progressive feature fusion is proposed,delivering exceptional performance in single image reconstruction.By integrating pixel attention mechanisms and large receptive field convolutions,two progressive pathways are devised to interpret inputs as features at varying levels of abstraction.This design enriches the network's capability to grasp contextual information while streamlining parameter redundancy.Through the utilization of hierarchical convolutions and multiple receptive field branches,the network sustains lightweight convolutions and acquires fused features from diverse scales on hierarchical pathways.This process enhances the network's proficiency in reconstructing edge details and intricate texture features.The experimental results reveal that the proposed algorithm achieves a peak signal-to-noise ratio of 32.47 dB on the Set5 benchmark test set and 28.81 dB on the Set14 test set.The algorithm surpasses existing advanced algorithms,utilizing fewer parameters and resulting in a 9％reduction in parameters,effectively validating the algorithm's efficacy.
image super resolutionattention mechanismreceptive field optimizationfeature fusion
NETL
NSTL
万方数据
