The work aims to propose a surface defect detection method of flat glass based on total reflection-grazing incidence combined lighting to solve the dust interferes with the detection of scratches and bright spots in defect detection of flat glass. By controlling the time sequences of lighting of total reflection and grazing incidence light sources, the glass images of glass in the corresponding time sequences of lighting were collected. The gray-scale, geometric characteristics and a series of relative deviation characteristics were calculated according to the difference of gray-scale texture of defects under different lighting. The BP neural network was developed to detect the dust and defects on the glass surface. In the end, the accuracy and recall rates of each category prediction of the BP neural network on the test set were all above 90%, and the overall accuracy rate reached 97.2%. From this point of view, the total reflection-grazing combined lighting imaging system has a simple structure, which reduces the difficulty of classification of dust and internal point defects in the glass image, and effectively reduces the misjudgment of dust and internal defects.