Face attribute recognition via end-to-end weakly supervised regional location

Facial attributes have been successfully applied in many fields of computer vision, such as face recognition, face retrieval, and face image synthesis. Locating attribute-related facial regions is a prerequisite for predicting the presence of attributes. However, most existing face attribute recognition methods obtain specific attribute regions through face segmentation annotations or face landmarks that are not easily available. In this paper, we propose a weakly supervised attribute location module (ALM) that can effectively detect facial regions with only image-level attribute labels and improve face attribute recognition using region-based local features. Moreover, we introduce a bottom-up skip connection structure to fuse the features from multiple convolutional layers, which can enhance attribute-specific region location with low-level spatial information supplements. Our network is easy to build and can be trained in an end-to-end manner. Extensive experiments demonstrate the superior performance of our method on LFWA and CelebA datasets.

Automated summary

In this paper, a weakly supervised attribute location module (ALM) is proposed to effectively detect facial regions with only image-level attribute labels, and improve face attribute recognition using region-based local features. Moreover, a bottom-up skip connection structure is introduced to enhance attribute-specific region location with low-level spatial information supplements. Extensive experiments demonstrate the superior performance of the proposed method on LFWA and CelebA datasets.