Semantic Frequency Prompt is a new concept for frequency domain analysis in knowledge distillation jointly proposed by Peking University, University of Sydney and Zhejiang University in November 2023.FreeKD: Knowledge Distillation via Semantic Frequency Prompt".

Semantic frequency cues determine the pixel imitation principle by interacting with the frequency bands, absorb the semantic frequency context when fine-tuning the teacher model, provide precise guidance for the student model to reconstruct the teacher frequency bands, and play a key role in knowledge distillation for dense prediction tasks. It aims to solve the limitations of traditional spatial domain methods through analysis and selective learning in the frequency domain, and provides a new perspective for knowledge distillation.

Its main functions include:

• Generate Pixels of Interest (PoIs) Masks: Generate pixel-level frequency masks to locate key pixels in different frequency bands by encoding the similarities between cues and frequency bands.

• Guiding the frequency domain learning of the student model: Through these masks, the student model can more accurately imitate the frequency characteristics of the teacher model, avoiding directly imitating redundant information in the low-frequency band or noise in the high-frequency band.