Facial Attribute Recognition Using Weighted Bottleneck Transformer for Intelligent Marketing

Facial attribute recognition (FAR) is essential in marketing strategies as it enables an intelligent robot to deliver relevant advertisements and attract more potential customers by profiling the audience. An efficient FAR is crucial for promising real-time operation on robots that use an edge device or low-cost single-board computer. This work proposes a compact CNN architecture and offers an efficient Weighted Bottleneck Transformer (WBT) block. WBT encourages the CNN architecture to capture the global representation by evaluating the interaction among image elements, previously weighted using the spatial attention mechanism. The proposed architecture, generating a few parameters and low computational operations, achieves competitive accuracy on two benchmark datasets: CelebA and LFWA. This architecture integrated with a face detector enables operating at 20.62 frames per second (FPS) on an Intel Core i7-97508 CPU in real-time using a captured face area image with a dimension of
$224\times 224$
. It confirms that the proposed recognizer enables implementation on a CPU device to support intelligent marketing.