GCA-ODN: A global context-aware dropout network for joint facial landmark detection and emotion recognition under occlusion.

Abstract

We propose the Global Context-Aware Dropout Network (GCA-ODN), a CNN-based, computationally practical neural architecture for joint facial landmark detection (FLD) and facial expression recognition (FER) under partial facial occlusion. GCA-ODN learns a shared embedding that encodes facial geometry and affective cues, improving reliability under face masks, large pose variations, and non-uniform illumination. The model integrates three complementary components: (i) a global context module that captures long-range spatial dependencies and injects holistic structural cues into local features, (ii) an attentive dropout module that generates attention-guided masks to suppress unreliable or overly dominant activations and reduce dependence on a limited set of visible cues, and (iii) a low-rank feature refinement module that promotes compact, noise-suppressed representations. Experiments across multiple FLD and FER benchmarks demonstrate competitive performance under the stated evaluation protocols, with clear improvements over the controlled ResNet-18 baseline and several occlusion-aware CNN methods. On 300W, GCA-ODN achieves NRMSE of 2.60×10-2 (Common), 3.20×10-2 (Full), and 5.61×10-2 (Challenging). For cross-dataset FER trained on AffectNet and evaluated directly on the target datasets without target-domain fine-tuning, it reaches 84.6% on CK+, 62.1% on JAFFE, and 68.3% on ISED. At an input resolution of 224 × 224, the model processes more than 80 frames per second with batch size one on an NVIDIA RTX A6000, demonstrating real-time throughput under the reported hardware configuration.