Category-level 6D pose estimation aims to identify the location and orientation of unseen objects. Extracting features from RGBD datasets and accurately reconstructing instance models are crucial to address challenges 6D pose estimation. In our research, we introduce an innovative three-branch parallel architecture with Attention Assistance Feature Fusion and Model Recursive Recovery network (AARR-Net), which comprises two essential components: (1) Firstly, we employ a three-branch parallel architecture network with channel attention assistance modules to extract weighted fused features from RGBD point clouds and images. (2) Secondly, combining the extracted shape priors of the categories, we produce a displacement map along with a correspondence matrix. An iterative recovery network is proposed to refine the parameters of the displacement map and correspondence matrix. Finally, we rebuild the 3D model in NOCS space and achieve 6D pose estimation. AARR-Net delivers results that are competitive with state-of-the-art methods based on shape priors on the CAMERA25 and REAL275 benchmark datasets.