Multimodal industrial anomaly detection is a challenging computer vision task. Recent unsupervised approaches rely on large-scale feature memory banks, which incur heavy storage and retrieval costs during inference and hinder their deployment in resource-constrained environments. Additionally, existing cross-modal interaction schemes lack an effective feature-fusion strategy, which makes it difficult to preserve geometric structure and texture details simultaneously, thereby leading to imprecise localization of anomalous regions. To address these issues, we propose an unsupervised multimodal industrial anomaly detection framework based on feature restoration, comprising feature extraction, feature alignment, feature reconstruction, cross-modal feature enhancement, and anomaly scoring modules. Within the framework, a dual-branch symmetric encoder–decoder network reconstructs RGB and point cloud features respectively, while self-attention and cross-attention modules further accomplish complementary cross-modal feature fusion. The entire framework is optimized with a joint reconstruction loss and a fusion consistency constraint. Extensive experiments on the MVTec 3D-AD and Eyecandies datasets show that our method outperforms previous state-of-the-art in 2D-3D multimodal industrial anomaly detection, thereby confirming its effectiveness and robustness in both intra-modal and cross-modal representation learning.