Abstract:  As technology continues to advance, the demand for accurately parsing complex real-world scenes is increasing. Although state-of-the-art semantic segmentation methods have achieved remarkable success in normal object detection, they are often challenged in glass surface detection. Glass has unique physical and chemical properties that make it lack its own distinct visual appearance and clear boundaries. Glass surfaces borrow the appearance of any scene or object behind them in real-world scenes, making the detection process difficult. Although some methods attempt to use additional training inputs or reference boundary information for glass surface detection, these methods are only effective with certain training requirements. Therefore, this paper aims to achieve effective glass surface detection without the need for additional training inputs or reference object information by utilizing the intrinsic properties of glass itself. This paper focuses on the image distortion caused by glass and designs a visual distortion aware module based on this phenomenon. The module extracts multi-scale visual distortion information and effectively integrates it to guide the backbone to focus on the areas of image distortion caused by glass surfaces, thereby locating the glass surfaces. In addition, this paper also notes that the structural information of glass surfaces helps to refine the prediction of glass masks, and utilizes the centroid information of glass surfaces in the form of a classification subtask to assist in predicting glass masks. Compared with state-of-the-art methods, this paper achieved significant improvements in standard evaluation metrics for the widely-used glass surface datasets.