Sketch face recognition based on light semantic Transformer network

Sketch face recognition has a wide range of applications in criminal investigation, but it remains challenging task due to the small-scale sample and semantic deficiencies caused by cross-modality differences. The authors propose light Transformer network extract model information images. First, employ meta-learning training strategy obtain task-related samples solve small problem. Then contradiction between high complexity problem sketch recognition, build transformer proposing hierarchical group linear transformation introducing parameter sharing, which can highly discriminative features on small–scale datasets. Finally, domain-adaptive focal loss reduce differences sketches photos improve effect network. Extensive experiments have shown that extracted proposed method significant effects. authors’ improves rate 7.6% UoM-SGFSv2 dataset, reaches 92.59% CUFSF dataset.