Efficient Self-Calibrated Convolution for Real-Time Image Super-Resolution

With the recent advancements in Convolutional Neural Net-work (CNN) architectures designs, many image processing tasks benefited from the design of such deep and complex networks including image super-resolution (SR). While deep and complex models achieve improved SR reconstruction, they lack the practicality in implementation for real-time applications, such as in mobile phones or online conferencing. This is due to the large number of parameters and excessive required multiply-accumulate operations (MACs). In this paper, an accurate real-time SR model structure is proposed. The proposed structure reduces the required number of MACs by performing all operations on low dimensional feature maps and reduces the model parameters by utilizing depthwise separable convolutional (DSC) layers. An efficient version of the recently introduced self-calibrated convolution with pixel attention (SC-PA) is introduced to further improve feature representation. Experimental results show that the proposed model improves performance in objective metrics, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) index, over similar complexity real-time SR models.