As the USV (Unmanned Surface Vehicles) works in offshore waters whose operating environment is complex, sonar images obtained are low-resolution, noisy and narrow overlap region, and ultimately led to the sonar image cannot be stitched or get a serious distortion after image mosaicing. In response to these problems, this paper proposes a sonar image mosaicing algorithm based on improved SIFT for USV. For sonar image registration, we applies RANSAC (RANdom Sample Consensus) algorithm to check feature points to improve the accuracy of image registration. For the problem that the distribution of original image brightness is different, this paper conducts brightness normalization algorithms for brightness adjustment. To avoid empty points in stitched image and the problem that mapping coordinates may not be an integer, this paper respectively take inverse mapping method and image interpolation algorithm to solve the problem. In order to improve the sonar image borders and edge fusion effect, this paper proposes a weighted fusion algorithm based on threshold and edge. Finally, the experiments show that the algorithm has a certain practicality.