The side-scan sonar towed working mode leads to changing local distortion in types and magnitude along the track, and further results in features' disposition and distortion in mosaic image. In order to guarantee clarity edge of the local features in overlap region of adjacent images, this paper proposes an elastic blocking mosaic method based on speeded-up robust features. Firstly, combining with the information of track line and swath, each stripe image can be sliced to several blocks. Then the SURF image registration is done in each block including features extraction, matching and refining. With the registration point-pairs between blocks from adjacent stripe image, a rigid transformation in global image and a series of elastic transformation in each block image can be done. The method not only eliminates systematic errors of position and heading between two adjacent side-scan sonar images, but also weakens random local distortion and achieves high-accurate registration in local area by corresponding features. At last, the experiment demonstrates the validity of this method, the overall registration accuracy reaches two pixels.