An Interactive Residual Fusion Network for Lightweight Super-Resolution Reconstruction

Abstract

Benefitting from attention mechanisms and deepening convolutional layers, CNN-based single image super-resolution reconstruction (SISR) methods evolve rapidly. However, computational cost, difficulty in global feature extraction and fixed convolutional kernel bottleneck the performance of these approaches. Therefore, a lightweight interactive residual fusion network (IRFN), which can realize adaptive kernel selection in channel and spatial dimension is proposed. Especially, by designing progressive enhanced large kernel attention (PELKA) in IRFN, global features and long-range dependence can be achieved effectively. Experimental results indicate that IRFN can achieve an excellent balance between performance and computational cost.