In order to overcome the difficulty that texture feature space dimension of multi-band infrared images is large and hard to analyze its imaging differences synthetically, a three-level feature selection model was proposed. Firstly tested consistency in the form of two contrast, and it was selected the features that difference distribution laws were the same. Secondly, the function of independent sample test based on SPSS software was used to test difference significant degree, and the features whose differences that were not significant were eliminated. Then, the correlation of features was analyzed, and one of the features stood for the features with high correlation. At the same time, the rest of features were reserved. Experiments prove that a few features selected by this model can reflect features differences of image, and it will provide references for target recognition and so on.