FNAT-Net: Feature space-based compression-aware adaptive thresholding network

In recent years, significant progress has been made in image compression sensing (ICS) through deep learning techniques. Deep Unfolding Networks (DUN) transforms the iterative reconfiguration process into an end-to-end deep neural network, improving interpretability and performance. However, traditional algorithms are limited to processing information in pixel space, missing the potential advantages of feature space. Additionally, most DUN are constrained by fixed input-output mirror structures that restrict information flow and lack adaptability due to their use of a fixed threshold for soft shrinkage operations. To address these limitations, we propose a novel feature space-based compression-aware adaptive threshold network (FNAT-Net). The supplementary information (FI) is utilized to enable FNAT-Net to perform fusion processing across both the pixel and feature domains, mapping a two-step approximate gradient descent algorithm from pixel to feature space. Furthermore, this paper introduces an effective enhanced Multi-Layer Perceptron (MLP) adaptive soft-thresholding strategy. This strategy enables FNAT-Net to address L1-regularized neighbourhood mappings with content-aware thresholds. FNAT-Net outperforms state-of-the-art methods, demonstrating superior performance across a wide range of scene changes and noise conditions.