Compact and progressive network for enhanced single image super-resolution—ComPrESRNet

The use of deep convolutional neural networks (CNNs) for single image super-resolution (SISR) in the recent years has led to numerous vision-based applications. Complementing the growing interest in the computer vision community embracing such networks, there is an unmet demand of reduced computational complexity. Despite being state-of-the-art for SISR tasks, CNN-based models need to be compact and efficient to account for applications running on low-cost deployment devices that have limited computation resources. While it is common to note that many state-of-the-art SISR approaches stack large number of convolutional layers in order to enhance their SR performance, there is proportional increase in the computational complexity. We propose a computationally efficient, compact and enhanced progressive network for SISR task which we hereafter refer as ComPrESRNet. The architectural changes employs a progressive learning strategy with a novel design of Enhanced densely connected parallel residual network (EDPRN) which simultaneously extracts rich features from the low-resolution (LR) observation while reducing the total number of parameters to 2.61M making it compact in nature that is suitable for low computational platform. The novelty of the proposed model stems from the inclusion of (i) densely connected ResBlock to extract rich features in the LR observation, (ii) extended global residual learning approach which stabilizes the training process effectively and also helps network to further improve the SR performance and (iii) progressive upscaling module which can generate an SR image of size $$\times 4$$ and $$\times 8$$ of original LR image. The robustness of the proposed method is further demonstrated on four different benchmark testing datasets consisting of natural scenes and urban landscape to exemplify the different applications. The superior performance over other state-of-the-art methods is also illustrated in this work for an upscaling factor $$\times 4$$ and $$\times 8$$ despite the lower computational complexity. The code of the paper is available at https://github.com/Vishal2188/Compactand-Progressive-Networkfor-Enhanced-SISR---ComPrESRNet .