Low-Resolution Face Recognition via Adaptable Instance-Relation Distillation

TL;DR

This work tackles low-resolution face recognition under data-distribution gaps between training and real-world testing. It introduces Adaptable Instance-Relation Distillation (AIRD), a teacher-student framework that jointly distills cross-resolution knowledge at the instance and relational levels, complemented by an online test-time adaptation stage called FaceBN. The method combines $\mathcal{L}_{IlD}$ and $\mathcal{L}_{RlD}$ losses to align embedding spaces while enhancing inter-class separation, and employs online batch-normalization statistics to bridge domain shifts without adding extra components. Experiments across verification and identification tasks demonstrate state-of-the-art performance and strong adaptability, validating AIRD’s effectiveness for practical, real-world LR-FR scenarios.

Abstract

Low-resolution face recognition is a challenging task due to the missing of informative details. Recent approaches based on knowledge distillation have proven that high-resolution clues can well guide low-resolution face recognition via proper knowledge transfer. However, due to the distribution difference between training and testing faces, the learned models often suffer from poor adaptability. To address that, we split the knowledge transfer process into distillation and adaptation steps, and propose an adaptable instance-relation distillation approach to facilitate low-resolution face recognition. In the approach, the student distills knowledge from high-resolution teacher in both instance level and relation level, providing sufficient cross-resolution knowledge transfer. Then, the learned student can be adaptable to recognize low-resolution faces with adaptive batch normalization in inference. In this manner, the capability of recovering missing details of familiar low-resolution faces can be effectively enhanced, leading to a better knowledge transfer. Extensive experiments on low-resolution face recognition clearly demonstrate the effectiveness and adaptability of our approach.

Figures (6)

• Figure 1: Our motivation. Direct knowledge transfer from high-resolution faces to low-resolution ones is hard due to resolution and data gaps. Thus, we conduct distillation in learning and perform adaptation in inference to enable sufficient and adaptable transfer.
• Figure 2: Overview of our adaptable instance-relation distillation (AIRD) approach. (a) The approach performs instance-relation distillation and then the learned student model can be used to effectively recognize low-resolution faces via inference adaptation. (b) Different from traditional knowledge distillation (KD), our AIRD can maximize the inter-class distance while minimizing the intra-class distance by combining instance-level distillation and relation-level distillation.$\mathcal{F}_{t}$ and $\mathcal{C}_{t}$ are the teacher backbone and head, while $\mathcal{F}_{s}$ and $\mathcal{C}_{s}$ are the student counterparts. $\mathbf{f}_{t}$ and $\mathbf{f}_{s}$ are the extracted features. $\mathbf{p}_{t}$ and $\mathbf{p}_{t}$ are the model predictions. $\mathbf{r}_{t}$ and $\mathbf{r}_{t,s}$ are the extracted feature relations.
• Figure 3: The score distributions achieved by different models on LFW.
• Figure 4: t-SNE feature plots on UCCS with four models.
• Figure 5: Data distribution statistics before (left) and after (right) FaceBN of LFW and CASIA-WebFace.