What question did this study set out to answer?

The aim is to develop a robust network for detecting facial landmarks and emotions even in challenging conditions with occlusions.

April 19, 2026Journal of King Saud University - Computer and Information Sciences1 citationsOpen Access

A context-aware dropout-based occlusion-adaptive network for robust facial landmark and emotion detection

MSMuhammad SadiqShenzhen University YZYunsheng ZhangShenzhen University YZYu ZhouKunming Institute of Botany

Key Points

The aim is to develop a robust network for detecting facial landmarks and emotions even in challenging conditions with occlusions.
Developed CAD-Net comprising three components: deep geometry-aware block, attentive dropout block, and low-rank learning block.
Utilized criss-cross attention to enhance global facial structure and correlation between various facial regions.
Implemented selective suppression of unreliable features through learnable dropout masks.
Conducted extensive experiments on standardized FLD and FER benchmarks.
CAD-Net demonstrated superior performance on standard benchmarks, particularly under severe occlusions.
The network maintains moderate computational overhead while improving robustness in facial analysis.
Empirical evaluations show the efficacy of CAD-Net across multiple datasets and applications.

Abstract

We present CAD-Net, a Context-Aware Dropout-based Occlusion-Adaptive Network for robust facial landmark detection (FLD) and facial expression recognition (FER) under challenging real-world conditions. Clinical and in-the-wild settings frequently involve strong occlusions (e.g., masks, medical devices), illumination changes, and large pose variations, where conventional convolutional neural networks (CNNs) often experience significant performance degradation. Rather than treating occlusion primarily as an isolated visibility-estimation problem, CAD-Net is designed to preserve structural facial consistency under partial visibility by jointly modeling long-range geometry, feature reliability, and compact landmark regression. CAD-Net comprises three complementary components: (i) a deep geometry-aware block that leverages criss-cross attention to preserve global facial structure and propagate information between spatially distant but correlated facial regions, (ii) an attentive dropout block that combines channel-wise attention with learnable dropout masks to down-weight unreliable or occluded regions, and (iii) a low-rank learning block that regularizes the regression head to obtain compact and stable landmark predictions. These modules are trained jointly within a unified framework that can be instantiated for both image-based and video-based facial analysis and naturally extended to a multi-task setting that couples FLD and FER. By integrating geometry-aware reasoning, selective occlusion suppression, and low-rank regularization in a single end-to-end architecture, CAD-Net improves robustness in a lightweight and practically deployable setting, while maintaining moderate computational overhead. Extensive experiments on standard FLD benchmarks (300W, COFW, AFLW, 300VW, Menpo) and FER datasets demonstrate that CAD-Net achieves competitive or superior performance compared with recent occlusion-aware methods, particularly under severe occlusions and pose variations. We further strengthen the empirical evaluation by reporting unified-protocol comparisons where feasible, as well as additional analyses of efficiency, stability across multiple runs, and cross-dataset transfer. The proposed design improves robustness without incurring prohibitive computational overhead, making CAD-Net suitable for time-sensitive biomedical and health informatics applications such as telemedicine, mental health monitoring, and elderly care.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper