Tomograms of thermal flow field are the information records characterized of fringes. Fringe analysis is a critical step both for interferometric and Moire tomography. How to extract clear and intact fringes from all kinds of noises, especially strongly inhomogeneous background illumination, is the key problem in preprocessing of thermal flow field tomogram. The characteristics of thermal flow field are analyzed and a probability filter method is developed, in which the center of image histogram is corrected, a so-called “extractable probability” is introduced as the standard for identifying signal from noise. According to a projection function determined by image features, gray image is projected into its probability plane where iterative signal enhancing and noise reduction are performed. It is proved that the probability filter presented in this paper is effective to the tomograms of thermal flow field.