An Adaptive Cross-Resolution Network for Low-Resolution Person Re-Identification

Person re-identification (PReID) plays a critical role in intelligent surveillance systems, yet its performance is often hindered by the low resolution of real-world CCTV imagery. The mismatch between degraded probe images and high-quality gallery images introduces a cross-resolution gap that significantly reduces recognition accuracy. Addressing this challenge requires models capable of recovering discriminative visual cues while maintaining consistent feature representations across heterogeneous input resolutions. To this end, we propose VarReID, a multi-resolution–adaptable hybrid framework that integrates image super-resolution with a robust PReID architecture to overcome the limitations of low-resolution PReID (LR-PReID). The framework incorporates a lightweight generative adversarial network (GAN)–based super-resolution module with a De-noise Enhancement Block to reconstruct essential structural and textural details from degraded inputs. The super-resolved images are then processed by a ResNet-50–based PReID network enhanced with Random Erasing, Linear Warm-Up, and Circle Loss to support stable and discriminative identity learning. A Resolution Adjustment Unit (RAU) further standardizes input dimensions, ensuring consistent feature extraction across varying resolutions. This integration of SR and PReID enables VarReID to handle low-resolution inputs more effectively than previous LR-PReID methods. Its stability is demonstrated in ablation studies, maintaining 89.30% Rank-1 at 32 × 64, 82.63% at 16 × 32, and 52.43% at 8 × 16, and achieving 77.99% on the naturally lowresolution VR-Market dataset. Compared with recent strong LR-PReID models such as RAPSR+RAReID (73.70% Rank-1), VarReID attains superior retrieval accuracy (77.99% Rank-1 and 89.24% Rank-5) due to clearer structural restoration and more resolution-consistent feature representations. VarReID was evaluated on DukeMTMCreID, VR-MSMT17, and VR-Market1501, achieving Rank-1 accuracies of 87.3%, 67.3%, and 77.99%, respectively, demonstrating its robustness under low-resolution conditions and its practical applicability.