Real-Time Facial Attribute Recognition Using Multi-Task Learning

In the realm of facial attribute recognition, crucial for applications like video surveillance, face retrieval, and recommendation systems, existing approaches often fall short in realistic scenarios, particularly for low-cost embedded systems. In this paper, we propose a Deep Multi-Task Learning approach to concurrently estimate multiple facial attributes from a single face image. We use convolutional neural networks to learn the commonalities and dissimilarities among various attributes. To address ordinal attribute estimation, we transform the original regression problem into a linear combination of binary classification subproblems, effectively reducing estimation errors. Experimental results from diverse datasets underscore the superior performance of our proposed approach. Finally, we present a practical solution for the cost-effective and swift application of our approach in realistic scenarios.