<p>Domain adaptive algorithms can enhance the performance of traditional object detection models in new scenarios. Although adversarial learning has achieved success in domain-adaptive object detection, it tends to align certain interference factors present in both the source and target domains, such as noise and blurry targets. As a result, the extracted domain-invariant features may incorporate these shared interference factors. Under the constraints of domain alignment, the model acquires domain-specific knowledge between the two domains, thereby enhancing the transferability of features. However, this process may compromise the discriminability of features, which can impact the accuracy of object detection. To address this, we propose a cross-domain object detector that incorporates multi-level domain feature refinement. First, we introduce a Feature Refinement Feed-forward Network (FRFN) in both the low-level and high-level parts of the backbone. This helps suppress irrelevant interference factors within the extracted domain-invariant features. Second, we design an instance-level domain-aware auxiliary classifier to enhance the model's discriminative ability while maintaining feature transferability. This classifier also learns domain-specific knowledge beneficial to object detection that might be discarded during domain alignment. Comprehensive experiments have shown that our method outperforms several popular cross-domain object detection methods in three challenging cross-domain scenarios: rainy days, cross-FOV scenes, and sim-real scenes.</p>