LiDAR technology provides cameras with precise depth perception; however, accurate registration between images and point clouds is essential to utilize this data. To address large-scale point cloud registration with images, an innovative pipeline leveraging semantic graph matching is presented. Recognizing the relative constancy of semantic information within scenes, semantic instances are extracted from both point clouds and images, constructing semantic graphs that reflect their 3D spatial relationships. Graph convolutional networks are subsequently employed to facilitate feature extraction and node matching within these semantic graphs. The experimental results show that the proposed method eliminates the need for initial pose estimation and is enhanced by joint optimization of camera intrinsic parameters.