AUTHOR=Shi Chaojun , Zhao Shiwei , Zhang Ke , Wang Yibo , Liang Longping 

TITLE=Face-based age estimation using improved Swin Transformer with attention-based convolution

JOURNAL=Frontiers in Neuroscience

VOLUME=Volume 17 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2023.1136934

DOI=10.3389/fnins.2023.1136934

ISSN=1662-453X

ABSTRACT=Recently, Transformer is a new direction in the computer vision field which is based on self multi-head attention mechanism. Compared with CNN, Transformer uses the self-attention mechanism to capture global contextual information and extract more strong features by learning the association relationship between different features, which has achieved good results in many vision tasks. In facial age estimation, some facial patches which contain rich age-specific information are critical in the age estimation task. However, Transformer ignores image specific properties, destroys the inherent internal structural information, and fails effectively in learning the age information of important facial patches. In this paper, we propose an attention-based convolution age estimation framework, called Improved Swin Transformer with attention-based convolution. In Improved Swin Transformer with attention-based convolution, two separate regions are implemented, namely the attention-based convolution (ABC) and Swin Transformer. ABC extracts facial patches that contain rich age-specific information by a shallow convolutional network and a multi-headed attention mechanism. Subsequently, the features obtained by ABC are spliced with the flattened image in Swin Transformer,  then input to Swin Transformer to predict the age of the image. ABC also introduced loss of diversity to guide training of self-attention mechanism, reducing overlap between patches so that the diverse and important patches are discovered. Through extensive experiments, we show that our proposed framework outperforms state-of-the-art methods on several age estimation benchmark datasets.