Frequency-decomposed attention joint optimization network for image compressive sensing

Abstract

Network approaches for image compressive sensing (ICS) have garnered significant attention due to their high efficiency and fidelity in image reconstruction. Reconstructing complex image textures from highly compressed measurements has been a longstanding goal of ICS, yet existing methods often struggle to varying degrees with the restoration of low-frequency (LF) textures and high-frequency (HF) details, which potentially limits the quality of the reconstructed image. In this paper, we propose a Frequency-decomposed Attention Joint Optimization Network (FAJO-Net) for ICS, which is capable of enhancing the attention to LF and HF components of images. Specifically, we introduce a frequency-decomposed sparse prior and coupling fidelity constraints, and incorporate a tri-optimization network framework for full, low, and high-frequency (FLH) features, where each component is optimized using an optimization-unfolded multi-scale network (OM-Net), inclusive of Principal Component Augmented Gradient Descent Module (PCAGDM) and U-shaped Proximal Mapping Module (UPMM). The PCAGDM optimizes the FLH features efficiently by supplementing the optimization of the minimum dimension principal component augmented features while optimizing the principal component features. The UPMM is able to perform multi-scale proximal mapping for all FLH features. Finally, we design a Frequency-decomposed Interaction Attention Module (FIAM) to enhance the fusion of FLH features, particularly the HF and LF components related to the full-frequency features, while reducing the impact of unnecessary features introduced by frequency decomposition. Extensive experiments demonstrate that our proposed FAJO-Net surpasses the state-of-the-art ICS networks in terms of image fidelity and visual effect, and validates that the proposed FAJO-Net framework can help enhance the image reconstruction capabilities of the vast majority of existing ICS networks, further unlocking the potential for high-fidelity restoration in ICS. Code is available at https://github.com/giant-pandada/FAJO-Net.