Human Facial Age Group Recognizer Using Assisted Bottleneck Transformer Encoder

Recognizing age from facial images has attracted considerable attention because of its wide array of applications and practical utilities. These include support for advertising platforms, access control, forensic objectives, and video surveillance. Efficient facial age recognition for these varied purposes is essential, necessitating smooth operation on low-cost devices or, at the very least, on a CPU to minimize implementation costs. This work proposes a lightweight CNN architecture efficiently integrated with a transformer encoder to perform facial age group recognition. An assisted bottleneck transformer encoder (ABTE) is introduced to enhance the feature extractor, generating only a few parameters and requiring low computation. As a result, the proposed architecture can achieve competitive performance on the two benchmark datasets, UTKFace and FG-NET. Moreover, this recognizer can attain real-time speed at 147 and 136 frames per second (FPS) with a single and double utilization of the ABTE, respectively, on a CPU device with Intel Core i7-9750H 2.6 GHz and 20 GB of RAM while maintaining its performance.