Reconstructing 3D structures from a single 2D image is a key challenge in computer vision, mainly due to the lack of depth information. Inspired by the human ability to understand object shapes through prior knowledge, we propose a method that integrates multi-shape priors and cross-modal features to improve reconstruction accuracy and completeness. Specifically, we build a compact latent space using a pretrained 3D autoencoder on ShapeNet point cloud data, obtain multiple structural prototypes via K-means clustering, and fuse them into a representative multi-shape prior point cloud using Chamfer Distance as weighting guidance. To enhance the image encoder’s focus on target regions, we design a Deformable Convolutional Attention Module (DCA-Module), which learns spatial offsets and modulation weights to effectively suppress background interference. To bridge the modality gap between images and point clouds, we further propose a 2D-3D feature fusion strategy that deeply integrates the visual appearance of images with the structural features of point clouds. Experimental results show that our method outperforms existing approaches on multiple benchmark datasets, achieving superior performance in structural completeness, geometric accuracy, and generalization capability.