AI-generated image is maliciously used in rumor propagation, which has highlighted the urgent need for robust AI image detection in standard and degraded scenarios. However, the performance of the existing detectors decreases significantly in the degraded scenario. This study systematically summarizes the various degraded scenarios encountered in AI image detection and proposes a novel two-stage detector architecture. We use only the real image dataset for training, the interference of degenerate artifacts to the detector is weakened by the Image Degradation Isolation Module (IDIM). The distortion artifacts generated in the photography process are extracted by the High-Frequency guided dual-stream positive Artifact Extraction Module (HF-AEM). Finally, refines the unified positive artifact representation in the real image by minimizing the distance from the features to the cluster center, which is used to determine the real image. The detector proposed in this study significantly outperforms existing detectors with 85.28% average accuracy in degraded scenarios, filling the research gap in AI image detection under degraded scenarios.