Efficient Dual-Branch Information Interaction Network for Lightweight Image Super-Resolution

Recently, deep convolutional neural networks (CNNs) have achieved remarkable success in single-image super-resolution (SISR) tasks. However, these methods often suffer from high computational and memory requirements, limiting their practicality for real-world applications. To address this challenge, we propose a lightweight and efficient dual-branch information interaction network (DIIN) for SISR. DIIN adopts a dual-branch structure that differs from the typical serial network architectures. Specifically, we design the CNN branch and Transformer branch as parallel structures. In the CNN branch, we employ a symmetric dual-branch feature interaction module (DFIM) to extract valuable local feature information. Concurrently, the Transformer branch utilizes a recursive Transformer to capture long-term global information and enhance reconstructed image details. By simultaneously considering these two branches, our model effectively combines the strengths of CNN in extracting local information and Transformer in capturing global information. Recognizing the complementarity of these two branches in SISR, we further incorporate a coefficient learning scheme to enhance their information interaction and obtain more comprehensive feature information, thereby improving overall model performance. Extensive experiments demonstrate that our DIIN outperforms competitive methods while consuming fewer computational resources and memory.

I. Introduction

Image super-resolution (SR) is a fundamental task to enhance low-resolution (LR) images by generating high-resolution (HR) counterparts, thereby improving their visual quality and capturing finer details. For more accurate analysis and measurement, SR is used to enhance the quality of images captured in instrumentation and measurement processes, leading to more precise and reliable results [1], [2]. However, this LR-to-HR mapping is ill-posed, as multiple HR images can be downsampled to yield the same LR image. In recent years, convolutional neural networks (CNNs) have gained significant attention in SR methods due to their exceptional feature extraction capabilities, surpassing traditional approaches [3], [4]. The groundbreaking work by Dong et al. [5] introduced the super-resolution CNN (SRCNN), which served as a foundation for subsequent CNN-based single-image SR (SISR) models. Building upon this, Kim et al. [6] proposed a highly deep SR model called very deep SR (VDSR), comprising 20 convolutional layers. VDSR exhibited superior performance compared to SRCNN, as the authors observed that deeper network architectures enable larger receptive fields. This capability allows the model to capture more contextual information and ultimately achieve better SR results.