This paper introduces a hybrid registration approach to build mosaics from side-scan sonar images. Due to specific acquisition procedure during surveys, standard SLAM techniques may not be robust enough to globally take into account a complete survey and correct sensor trajectories in order to properly georeference every pixel from all these images. iSAM algorithm has been fed with real side-scan images and shows interesting capabilities to produce corrected sensor trajectories allowing relevant coarse image registration, based on landmarks extraction and pairing. These trajectories will then guide a block-matching procedure that will refine these trajectories by finely matching only sonar images relevant areas.