Computationally Efficient Super-Resolution Approach for Real-World Images

Abstract

AbstractMost of the existing single image super-resolution (SISR) methods are trained and evaluated on synthetic datasets in which the low-resolution (LR) images are synthesized with simple and uniform bicubic degradation. Those methods perform better on a synthesized testing dataset but fail to obtain better super-resolution (SR) results on real-world images. However, by stacking more convolution layer, the SR performance can be improved. But, such techniques increase the number of training parameters and hence offer heavy computational burden on resources which make them unsuitable for real-world applications. To solve this problem, we propose a computationally efficient SR approach called real-world super-resolution network (RSRN) to super-resolve the real-world images. In RSRN, we propose a novel residual block called densely connected parallel residual block (DPRB) which helps to extract more complex features of LR observations. To prove the effectiveness of the proposed RSRN method, we train the proposed model on real-world images as well as on synthetic dataset. We compare the SR performance of the proposed method with that of other existing SISR methods and observe that the proposed RSRN method obtains better SR performance with more high-frequency details than that of the recently proposed state-of-the-art SISR methods with significantly less number of training parameters.KeywordsSuper-resolutionReal-world imageConvolutional neural networkDensely connected parallel residual blocks