LDR-Net: Landmark-Guided Diverse Regional Representation Learning for Facial Expression Recognition

Reliable facial expression recognition (FER) is essential for human-computer interaction across scenarios. However, existing methods often overlook spatial structure when seeking region-specific features due to fixed or limited local regions, limiting representation capability under real-world variations. Inspired by the multi-location attention of the human visual system, we propose a Landmark-guided Diverse Regional Representation Network (LDR-Net), using dynamic landmarks as patch centers to preserve structural details while locating expression-critical areas, facilitating effective regional representations for FER. First, a novel Diverse Regional Feature Extraction (DRFE) module operates via complementary operations: landmark-guided cropping for local details and cross-level integration with feature reorganization for holistic aggregation. Second, a novel Diverse Representation Learning (DRL) module is proposed with a collaborative dual-stream mechanism that captures fine-grained local dependencies via Transformers while reinforcing global features through attention-based enhancement, enabling comprehensive feature learning. Finally, a new Hybrid Feature Fusion (HFF) module is proposed for joint decision optimization via a hierarchical hybrid strategy, which aggregates intra-branch predictions followed by weighted branch-level fusion. Experiments on three FER benchmarks (RAF-DB, AffectNet, SFEW) and five occlusion/pose test sets demonstrate that LDR-Net outperforms state-of-the-art methods, while cross-scene validation on KMU-FED confirms its effectiveness in real-world driving.