Texture plays an important role in image analysis. There are mainly two categories of texture analysis techniquesstatistical methods and structural methods, both of which have their own advantages. Based on the two categories, a novel approach is proposed, which describes linear texture using a kind of ARG (Attributed Relational Graphs), and is applied to image retrieval using probabilistic relaxation algorithm. Because the image 's characteristics are presented intuitively through structural attributes of graph, this method has excellent ability of texture description. Furthermore, straight-line segment is extracted as texture primitive. Three geometrical relation attributes between texture primitives, which are invariant to changes of scale, rotation and translation, are put forward. Therefore the applications of image retrieval basing this method possess excellent anti-noise ability. Practical experiments show that the performance of the new approach is satisfying.