Single Image Super-Resolution Using ConvNeXt

In recent years, a lot of deep convolution neural networks have been successfully applied in single image super-resolution (SISR). Even in the case of using small convolution kernel, those methods still require large number of parameters and computation. To tackle the problem above, we propose a novel framework to extract features more efficiently. Inspired by the idea of deep separable convolution, we improve the standard residual block and propose the inverted bottleneck block (IBNB). The IBNB replaces the small-sized convolution kernel with the large-sized convolution kernel without introducing additional computation. The proposed IBNB proves that large kernel size convolution is available for SISR. Comprehensive experiments demonstrate that our method surpasses most methods by up to 0.10 ~ 0.32dB in quantitative metrics with fewer parameters.