The manufacturing process of TFT-LCD is comprised of about 100 working procedures. Though most of them are performed in net rooms, TFT-LCD panels would exhibit some visual defects inescapability. Mura defect is a familiar kind of region defect. Mura is a Japanese word meaning blemish that has been widely adopted by the display industry to provide a name for low contrast, non-uniform brightness regions, typically larger than single pixels. Moreover, mura is the biggest challenge to inspect among the defects of TFT-LCD. The main reasons are as follows: 1) While inspecting mura, an inspector need to observe the display panel from a slantwise angle of view. In this case, TFT-LCD panel shows an uneven brightness due to the variance of the backlight and uneven distributions of liquid crystal material. Moreover, mura defects generally have low contrast and vague outline. Therefore, it is very difficult to segment mura area from TFT-LCD image effectively. 2) The estimation of mura is subject to the nature and characteristic of mura as well as the visual mentality of human obsever. Therefore, it is difficult to accomplish the task of mura inspection perfectly with traditional methods which merly employ the contrast and area of object as the features of mura. 3) The nature of mura and its criterion is of complexity and vagueness. Therefore, traditional pattern recognition based on classical set theory can not perform mura inspection imitateing hurman observer. Consequently, automatic mura inspection of TFT-LCD emphasizes on the image segmentation of mura under complex background, feature extraction in accordance with mura and pattern recognition method being capable of imitating human beings. A newMurafpralgorithm for visual defect inspecting of TFT-LCD is described in this paper.In image segmentation, a segmentation method of mura image based on polynomial surface fitting by least square method is proposed in this paper. Surposed that the gray value of each pixel of a digital iamge is the function of its two-dimensional coordinates and constitutes a set of spatial data distributed on rectangular grids together with the latter. Then a polynomial surface fitting is performed on the spatial data, that is to say a polynomial surface is solved to fit