Low-Resolution Object Recognition with Cross-Resolution Relational Contrastive Distillation

TL;DR

This study proposes a cross-resolution relational contrastive distillation approach to facilitate low-resolution object recognition that enables the student model to mimic the behavior of a well-trained teacher model which delivers high accuracy in identifying high-resolution objects.

Abstract

Recognizing objects in low-resolution images is a challenging task due to the lack of informative details. Recent studies have shown that knowledge distillation approaches can effectively transfer knowledge from a high-resolution teacher model to a low-resolution student model by aligning cross-resolution representations. However, these approaches still face limitations in adapting to the situation where the recognized objects exhibit significant representation discrepancies between training and testing images. In this study, we propose a cross-resolution relational contrastive distillation approach to facilitate low-resolution object recognition. Our approach enables the student model to mimic the behavior of a well-trained teacher model which delivers high accuracy in identifying high-resolution objects. To extract sufficient knowledge, the student learning is supervised with contrastive relational distillation loss, which preserves the similarities in various relational structures in contrastive representation space. In this manner, the capability of recovering missing details of familiar low-resolution objects can be effectively enhanced, leading to a better knowledge transfer. Extensive experiments on low-resolution object classification and low-resolution face recognition clearly demonstrate the effectiveness and adaptability of our approach.

Figures (5)

• Figure 1: A human who is more familiar with a high-resolution object can recognize the corresponding low-resolution one better. Transferring the structural relation knowledge between different resolution samples can help recognizing low-resolution objects. Our cross-resolution relational contrastive distillation enables low-resolution samples ($f^{S_i}$) to mimic the structural relation between corresponding high-resolution sample ($f^{T_i}$) and other high-resolution samples ($f^{T_j}, i\neq j$).
• Figure 2: The framework of our approach. The approach performs knowledge transfer from high-resolution teacher to low-resolution student by sufficiently modeling high-order representation relations, which simultaneously addresses knowledge distillation and low-resolution recognition in a single framework.
• Figure 3: Face verification accuracy on LFW under different negative number (left) and distillation temperature (right).
• Figure 4: t-SNE feature plots by baseline (left) trained with softmax loss and CRRCD (right) on SVHN.
• Figure 5: The distribution of cosine similarity score under low-resolution setting on LFW by ArcFace (left) and CRRCD (right). The x-axis represents the cosine similarity of face pairs, and y-axis is the frequency. The negative pairs and positive pairs are marked in blue and orange, respectively.