Image restoration (IR) focuses on enhancing degraded images to reconstruct their original high-quality form and is fundamental to many vision applications. Although deep learning (DL) has recently shown impressive results in IR, its inherent black-box characteristics often hinder interpretability and transparency. To tackle this issue, we introduce an innovative deep unfolding framework derived from an optimization model based on image decomposition. The IR task is reformulated into two subtasks: residual layer ( \({\mathcal{R}}\) ) reconstruction and background layer ( \({\mathcal{B}}\) ) recovery. Using Proximal Gradient Descent (PGD), we design an iterative algorithm that unfolds into a neural network, where each module corresponds to an iteration step, ensuring strong interpretability. In addition, we design a Multi-Scale Channel Attention (MSCA) module that strengthens the extraction of multi-scale features and adaptively adjusts the significance of different channels. Comprehensive evaluations across various IR tasks show that our approach delivers competitive results with strong interpretability.