Abstract: This paper proposed an algorithm to extract texture primitive dynamically according to image itself and then get features from texture primitive to retrieval images. Firstly, do LBP transform on images, and set LBP value cycle as texture primitive initial size; then, adjust texture primitive window pixel by pixel to get final texture primitive size; last, calculate image content similarity by features extracting from texture primitive. Experimental results show that the proposed retrieval method has great performance compared with the existing method which is based on whole images in retrieving images with repetitive texture primitive.