Abstract
Recently, deep convolutional neural networks (CNNs) have been successfully applied to the single-image super-resolution (SISR) task with great improvement in terms of both peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). However, most of the existing CNN-based SR models require high computing power, which considerably limits their real-world applications. In addition, most CNN-based methods rarely explore the intermediate features that are helpful for final image recovery. To address these issues, in this article, we propose a dense lightweight network, called MADNet, for stronger multiscale feature expression and feature correlation learning. Specifically, a residual multiscale module with an attention mechanism (RMAM) is developed to enhance the informative multiscale feature representation ability. Furthermore, we present a dual residual-path block (DRPB) that utilizes the hierarchical features from original low-resolution images. To take advantage of the multilevel features, dense connections are employed among blocks. The comparative results demonstrate the superior performance of our MADNet model while employing considerably fewer multiadds and parameters.
Introduction
Single-Image Super Resolution (SISR) involves reconstructing a high-resolution (HR) image from a single low-resolution (LR) input. While deep learning has significantly improved reconstruction quality, state-of-the-art models often rely on massive architectures with millions of parameters. This high computational demand makes them impractical for resource-constrained environments like mobile devices or real-time video processing.
MADNet (Multiscale Attention Dense Network) addresses these challenges by focusing on computational efficiency without sacrificing visual quality. The architecture is designed to exploit hierarchical features—information embedded in intermediate layers—which are often underutilized in traditional models.
Proposed Methodology
The MADNet architecture is divided into two primary stages: the Efficient Feature-Extraction Network (EFEN) and the Upsampling Network (UN).
The core innovations within this framework include:
- Residual Multiscale Module with Attention Mechanism (RMAM): To capture features at various granularities, MADNet utilizes parallel branches with differing kernel sizes and dilation rates. This multiscale approach is augmented by a Squeeze-and-Excitation (SE) block, which provides channel-wise attention to prioritize the most relevant features.
- Dual Residual-Path Block (DRPB): This structure facilitates better gradient flow and information preservation. By allowing low-frequency information to bypass certain layers, the network can dedicate more capacity to learning complex high-frequency details like textures and edges.
- Dense Connections: By implementing dense connections between blocks, MADNet ensures that features from early stages are reused, mitigating the vanishing gradient problem and maximizing the utility of hierarchical data.
Experimental Results
The model was trained on the DIV2K dataset and evaluated against standard benchmarks including Set5, Set14, B100, and Urban100.
Efficiency and Performance
MADNet demonstrates a superior trade-off between accuracy and computational cost. For instance, at a $3\times$ upscaling factor on the Urban100 dataset, MADNet achieved a PSNR of 28.21 dB while requiring only 88.4G Multi-Adds. This is significantly more efficient than models like VDSR (612.6G) or DRRN (6,796.9G).
Qualitative Analysis
Visual comparisons reveal that MADNet excels at recovering fine, repetitive textures—such as architectural details—where larger models often produce blurry artifacts. Ablation studies further highlighted the importance of the attention mechanism, which contributed an improvement of approximately 0.8 to 0.9 dB in PSNR. Additionally, the use of a custom $L_F$ loss (combining $L_1$ loss with Total Variation regularization) resulted in sharper, more realistic image edges.
Conclusion
MADNet provides a practical solution for high-quality SISR on edge devices. By combining multiscale feature extraction with dense connections and attention mechanisms, it achieves competitive performance with a fraction of the computational overhead of traditional deep models. Future research will focus on extending this architecture to "blind" super-resolution and video-based tasks where temporal consistency is paramount.
Summary updated