Hybrid CNN-SSM Network for Efficient and Perceptually Enhanced Super-Resolution

This paper introduces HybSR, a novel image super- resolution method that leverages the strengths of both Convolutional Neural Networks (CNNs) and State Space Models (SSMs) within a hybrid architecture. CNNs effectively capture local image details, while SSMs offer linear scalability and excel at modeling long-range dependencies. By combining these two paradigms, HybSR achieves a compelling balance between efficiency and performance. Trained with a hybrid loss function encompassing pixel-wise and perceptual components, the model ensures both accurate reconstruction and high visual fidelity. Furthermore, knowledge distillation from a larger teacher network to a smaller student network enhances the model's efficiency. Ex- tensive evaluations on benchmark datasets demonstrate HybSR's superior performance in terms of reconstruction quality, model size, and computational efficiency compared to existing efficient super-resolution methods.

I. Introduction

The growing demand for high-resolution visuals in fields like medical imaging, surveillance, and entertainment has driven significant research in image super-resolution (SR), which reconstructs high-resolution (HR) images from low- resolution (LR) inputs, addressing issues like pixelation and degradation. Recent advancements in deep learning, particularly Convolutional Neural Networks (CNNs), have revolutionized single image super-resolution (SISR) by leveraging CNNs' ability to learn hierarchical representations and capture local patterns. However, SR deployment faces challenges due to computational and memory constraints, especially in resource-limited environments, prompting the development of efficient SR methods capable of high-quality reconstructions under such conditions. Additionally, traditional metrics like Peak Signal-to-Noise Ratio (PSNR) may not fully capture human-perceived image quality, which emphasizes natural- ness, texture realism, and visual appeal. Addressing these challenges, this paper introduces HybSR, a hybrid network combining CNNs and State Space Models (SSMs) to balance efficiency, accuracy, and perceptual quality in SR.