Introduction
Facial Expression Recognition (FER) is a critical component of human-centered computing, yet it faces significant hurdles in balancing global facial context with fine-grained local details. Traditional approaches often rely on manual annotations for facial parts or fail to capture the complex relationships between specific features like the eyes and mouth. To address these challenges, researchers have introduced the Weakly Supervised Local-Global Relation Network (WS-LGRN). This model autonomously identifies key facial regions and intelligently integrates them with global information to achieve superior recognition performance.
The WS-LGRN Architecture
The WS-LGRN framework operates through a sophisticated multi-stage pipeline designed to mimic the human visual system's ability to focus on both the "big picture" and specific details.
As shown in the architecture diagram, the process is divided into two main stages:
- Stage 1: Weakly Supervised Learning. An Attention Map Generator (AMG) is pre-trained on the CelebA dataset. By using image-level attribute labels (e.g., "smiling" or "bushy eyebrows"), the AMG learns to localize expressive facial regions without requiring expensive pixel-level masks.
- Stage 2: Relation Learning. The parameters of the AMG are transferred and fixed. The network then extracts Local Feature Embeddings (LFE) and Global Feature Embeddings (GFE). These are processed by a Relational Reasoning Unit (RRU) before final classification.
Weakly Supervised Attention
The core innovation of WS-LGRN lies in its use of "weak supervision" to solve the localization problem. By leveraging existing facial attribute datasets, the model gains a prior understanding of where eyes and mouths are located.
The generated attention maps serve as a guide for Bilinear Attention Pooling, which filters out background noise and focuses the network's representational power on the most informative parts of the face. This allows the model to be robust across different identities and lighting conditions.
Relational Reasoning and Contrastive Loss
Rather than simply concatenating local and global features, the Relational Reasoning Unit (RRU) employs a "Fuse and Reasoning" strategy. It adaptively weighs the importance of different regions based on the specific input image. For instance, it might prioritize the mouth for a "happy" expression while focusing more on the eyes for "fear."
To further enhance discriminative power, the authors implemented a Multi-Granularity Contrastive Loss. This loss function operates at both global and local levels, ensuring that images of the same expression are clustered together in the feature space while different expressions are clearly separated. This effectively reduces intra-class variation caused by individual identity differences.
Experimental Results
The WS-LGRN was evaluated on two benchmark datasets, yielding state-of-the-art results:
- CK+ (Cohn-Kanade): The model achieved an accuracy of 98.37%, surpassing previous leading methods like FMPN and DeRL.
- RAF-DB (Real-world Affective Face Database): In a more challenging real-world setting, the model reached 85.20% accuracy, demonstrating its robustness against the variability found in unconstrained internet images.
Ablation studies confirmed that the RRU's adaptive fusion and the multi-granularity contrastive loss were both essential components in reaching these performance levels.
Conclusion
The WS-LGRN represents a significant step forward in FER by demonstrating that facial attributes can provide effective "weak" supervision for localizing expressive regions. By combining this localization with adaptive relational reasoning and multi-level contrastive learning, the network achieves a nuanced understanding of facial expressions that rivals human perception. Future work may extend this weakly supervised approach to other facial analysis tasks, such as face alignment and attribute manipulation.
Summary updated