<p>Deep learning-based single-channel blind image separation methods often treat images as regular grids in Euclidean space and employ fixed-size convolutional kernels when processing them, which limits their ability to model nonlinear structures, such as irregular objects within images, and to capture spatial dependencies between pixels. To address this issue, We propose novel graph-structured mathematical models for single-channel blind image separation (SCBIS) and introduce a lightweight integrated algorithm that fuses both global and local features by combining Graph Isomorphism Networks (GIN) and Chebyshev Graph Convolutional Networks (ChebNet). The graph structure is introduced to more accurately characterize the spatial dependencies among image pixels. Meanwhile, the message-passing mechanisms of GIN and ChebNet are leveraged to enable dual-path parallel transmission of global and local information. Finally, a multi-head attention mechanism is employed to aggregate global and local features. Experimental evaluation on several benchmark datasets demonstrates the superiority of our approach over state-of-the-art methods, as it effectively models the overall structure during the separation process, obtains a more accurate unmixing matrix, improves separation performance, and reduces model complexity.</p>