<p>Compressive sensing (CS) is a signal processing paradigm that enables acquisition and compression at the same time. Deep Unfolding Network (DUN) is a widely used reconstruction technique for deep learning-based compressive sensing. By unfolding traditional algorithms into cascaded neural networks, DUN achieves good interpretability. However, most existing methods treat image patches uniformly in the reconstruction stage and overlook texture complexity. Besides, existing Transformer-based methods calculate correlations between all tokens even if they are weakly related. To address these issues, we propose a novel saliency guided deep unfolding network (SGDUN), which divides image patches and reconstructs them in groups according to their textures. With our designed saliency map, SGDUN can dynamically choose the appropriate reconstruction module for each patch and perform attention mechanism within groups of similar texture complexity. Experimental results demonstrate that our method provides excellent scores on benchmark datasets and achieves better reconstruction performance than state-of-the-art deep learning-based CS methods.</p>