FAR-AMTN: Attention Multi-Task Network for Face Attribute Recognition

TL;DR

FAR-AMTN tackles the scalability and generalization challenges of multi-task face attribute recognition by introducing Weight-Shared Group-Specific Attention (WSGSA) to efficiently learn group features, Cross-Group Feature Fusion (CGFF) to enable interactions among attribute groups, and a Dynamic Weighting Strategy (DWS) to synchronize task convergence. With a ResNet50 backbone, FAR-AMTN achieves state-of-the-art accuracy on CelebA and LFWA while markedly reducing parameter count and memory usage, outpacing existing MTN approaches. The CGFF module captures nonlinear relations across attribute groups, the WSGSA module provides parameter-efficient, group-specific attention, and DWS balances gradient speeds and loss scales during training. Together these components yield superior performance and efficiency, validated by ablations that isolate each module’s contribution and by strong comparative results on benchmark FAR datasets.

Abstract

To enhance the generalization performance of Multi-Task Networks (MTN) in Face Attribute Recognition (FAR), it is crucial to share relevant information across multiple related prediction tasks effectively. Traditional MTN methods create shared low-level modules and distinct high-level modules, causing an exponential increase in model parameters with the addition of tasks. This approach also limits feature interaction at the high level, hindering the exploration of semantic relations among attributes, thereby affecting generalization negatively. In response, this study introduces FAR-AMTN, a novel Attention Multi-Task Network for FAR. It incorporates a Weight-Shared Group-Specific Attention (WSGSA) module with shared parameters to minimize complexity while improving group feature representation. Furthermore, a Cross-Group Feature Fusion (CGFF) module is utilized to foster interactions between attribute groups, enhancing feature learning. A Dynamic Weighting Strategy (DWS) is also introduced for synchronized task convergence. Experiments on the CelebA and LFWA datasets demonstrate that the proposed FAR-AMTN demonstrates superior accuracy with significantly fewer parameters compared to existing models.

Figures (9)

• Figure 1: (a) The preceding methodology of multi-task attribute recognition. (b) The FAR-AMTN method. To address three issues in existing methods: the increase in model parameters with the number of tasks, the heterogeneity of group features, and asynchronous convergence, we propose specific group attention layers with parameter sharing mechanism and cross-group feature fusion modules. Additionally, we enhance the dynamic weighting $\lambda$ strategy to enable synchronous convergence for multiple tasks. Where $\otimes$ denotes the element-wise multiplication of two elements.
• Figure 2: FAR-AMTN architecture diagram. The proposed FAR-AMTN architecture consists of Shared Bottom module, Weight-Shared Group-Specific Attention (WSGSA) module, and Cross-Group Feature Fusion (CGFF) module. In the diagram, $\otimes$ represents the element-wise multiplication of two elements. The dynamic weighting strategy is in the section \ref{['sec:Dynamic Weighting Strategy']}.
• Figure 3: Cross-group feature fusion visualization.
• Figure 4: Comparing the train and test losses of Dynamic Weighting Average and Dynamic Weighting Strategy.
• Figure 5: Comparison of experimental results between FAR-AMTN and the latest method. On the CelebA and LFWA datasets, respectively.