Texture images of limited size can be insufficient for statistical learning to perform the task of image retrieval. This paper proposes a hypothesis that the utilization of synthesized texture and noise addition can improve texture analysis with spatial statistics. Rationales for this hypothesis are that texture synthesis allows the enlargement of an image size for better description of its spatial statistics, and noise added to texture pixels at some levels enhances discriminative power of texture randomness. The improvements using synthesized images of coarse-aperiodic and fine-periodic texture categories with added noise for feature extraction with the semi-variogram and feature-vector matching using the log-likelihood ratio suggest the validation of the proposed hypothesis that is promising for handling classification of texture of small sample sizes.