A hybrid Mamba–Transformer model for image restoration

Image restoration plays a critical role in many engineering systems, including remote sensing, medical imaging, and industrial inspection, where degraded visual data can significantly impair downstream analysis and decision-making. However, existing restoration methods often struggle to balance restoration quality, computational efficiency, and model interpretability, which limits their practical deployment in real-world engineering scenarios. In this paper, we propose Mamba–Transformer for Image Restoration (MatIR), a hybrid image restoration framework that integrates transformer-based (Transformer) attention with state–space modeling (SSM) in a knowledge-guided and modular architecture. MatIR introduces a Channel-Triangle Attention (CTA) module to jointly model global semantic consistency and local structural integrity, and an Image Restoration State Space (IRSS) module that efficiently captures long-range spatial dependencies using multi-directional state–space scanning with linear complexity. A serial stacking strategy is adopted to progressively refine features while avoiding redundant computation. Extensive experiments on multiple image restoration tasks, including denoising, deblurring, and super-resolution, demonstrate that MatIR consistently outperforms state-of-the-art methods, achieving up to 0.3–0.6 decibels (dB) peak signal-to-noise ratio (PSNR) improvements and lower learned perceptual image patch similarity (LPIPS) scores across benchmarks. These results indicate that MatIR delivers both high-fidelity reconstruction and improved perceptual quality. Overall, MatIR provides a scalable, interpretable, and efficient solution for artificial intelligence (AI)-driven image restoration, offering clear potential for deployment in practical engineering applications.