Journal Article

SCCADC-SR: a real image super-resolution based on self-calibration convolution and adaptive dense connection

Yang, Xin, Li, Hengrui, Li, Tao, Wu, Chenhuan

2023

Overview

Because the real degradation model is more complex, and the different computing performance of devices leads to different degradation results. The super-resolution based on the real image degradation model has great challenges in practical applications. To solve these problems, we propose a novel SR network based on self-calibration convolution and adaptive dense connection (SCCADC-SR). Firstly, we introduce self-calibration convolution as the basic convolution module and use it as a supplement to the attention mechanism. Secondly, we use efficient channel attention (ECA) to construct an adaptive dense connection structure to deal with the features at the different levels. Then, we use the CutBlur method to enhance the data to improve the generalization ability of the model and use the long skip connection to improve the convergence of the depth model structure. Finally, SCCADC-SR combines self-ensemble and model ensemble to improve the model’s robustness and reduce the noise. Experimental results show that for both real image data and Bicubic data, our SCCADC-SR improves SR reconstruction performance by 5% compared with the state-of-the-art methods.

• Publisher: Springer Nature B.V
• Subject: Calibration / Convolution / Image degradation / Image resolution / Self calibration