SSL-LODDA: Self-supervised learning for low-light object detection with domain adaptation

Object detection under low-light conditions is severely challenged by illumination degradation, noise, and large distribution discrepancies between well-lit source domains and low-light target domains. Moreover, the lack of annotated low-light data significantly limits the effectiveness of supervised learning approaches. To address these issues, this paper presents a unified self-supervised domain-adaptive object detection framework that jointly integrates self-supervised representation learning, adversarial and statistical domain alignment, GAN-based synthetic augmentation, and pseudo-label refinement within a multi-loss optimization scheme. The proposed method first employs contrastive self-supervised learning to pretrain a shared feature extractor using unlabeled data, enabling the learning of illumination-invariant and transferable representations. To reduce cross-domain feature discrepancies, Domain-Adversarial Neural Networks (DANN) are applied to enforce domain confusion through adversarial training, while Maximum Mean Discrepancy (MMD) explicitly minimizes higher-order distribution shifts between source and target feature spaces. In addition, a GAN-based low-light image synthesis module augments the target domain with diverse illumination degradations, improving robustness under extreme lighting conditions. Pseudo-labeling is further introduced to exploit high confidence predictions on unlabeled target samples for supervised fine-tuning. All components are optimized jointly using a unified loss function that combines detection, self-supervised, DANN and MMD objectives. Extensive experiments on PASCAL VOC, Cityscapes, MS COCO and ExDark under multiple cross-domain adaptation settings demonstrate that the proposed framework consistently outperforms state-of-the-art domain-adaptive detectors. The method achieves up to 82.3% mAP on challenging low-light adaptation tasks, while ablation studies confirm the complementary contributions of SSL, DANN, MMD, and pseudo-labeling. These results validate the effectiveness of the proposed approach for scalable and robust low-light object detection in real-world applications.