A Cluster Tree Network for Image Super-Resolution

Image super-resolution (SR) enhances the clarity and detail of images produced by consumer electronics, particularly in high-end devices, offering a superior visual experience. In this paper, a SR technique is proposed to enable effective and reliable processing in consumer electronics. Deep networks extract hierarchical structural feature through multiple layers for improving image SR results, however, their robustness may be limited in complex scenes. To address this, a cluster tree network for image SR (CTSRNet) was developed. The proposed tree architecture leverages hierarchical information by branching feature extraction into pathways that process varying, levels of abstraction and complexity, unlike monolithic convolutional neural networks (CNNs), where deep features may suppress shallower ones. Adaptability to diverse scenes is enhanced through conditional parameterized convolutions (Condconv), which learn salient features. By clustering image regions based on their properties, the networks learns inter-region to recover finer details. Experimental results demonstrate that CTSRNet surpasses popular SR methods. With reduced complexity and resource demand, the proposed approach is well-suited for edge computing in consumer electronics. The code is available at https://github.com/xwq325/CTSRNet.