Objective To further improve the quality of document image text detection, in-depth research was performed to analyze how statistical features influenced the classification of text texture. Methods First, the document images’feature images were obtained through Gabor-statistical feature, and then the SCA algorithm was applied to extract the text and non-text samples. Finally, SVM was employed to fulfill the text detection. To choose the statistical feature, Fisher criteria were used. Results The experiments implied that homogeneity returned the maximum class separation distance according to Fisher criteria and gave the best detection result. Conclusion A relatively good detection result could be obtained using Gabor-homogeneity feature when dealing with different types of document images.