Eskn: Enhanced Selective Kernel Network For Single Image Super-Resolution

For single image super-resolution (SISR), one recent research direction is to build an effective multi-scale context extraction pipeline via parallel convolutional streams. Although very competitive SR performance has been achieved, effective solutions for extracting and integrating multi-scale context are still underexplored. We propose an enhanced selective kernel module (ESKM) to address this challenging problem and build a network that achieves high-quality SISR. The key of the proposed ESKM is to perform self-learned filter-oriented weights re-calibration to better extract insignificant but important features which are critical for high-accuracy SISR tasks. Moreover, we replace the Softmax operation with Sigrnoid for more flexible weights learning and remove the dimension reduction/expansion component to build a direct correspondence between channels and their weights. We also design a symmetric connection scheme (SCS) to better fuse the hierarchical features extracted from different convolutional stages. More specifically, the low-level features are adjusted via a spatial attention module to achieve more effective fusion with high-level semantic features. We then stack multiple ESKMs via SCS to build our new network, named Enhanced Selective Kernel Network (ESKN). Extensive experimental results demonstrate the effectiveness of our proposed ESKN model, outperforming the state-of-the-art SISR methods in terms of restoration quality and network complexity. (C) 2021 Elsevier B.V. All rights reserved.