This paper proposes a novel network named ArtifactScope for generated image detection, aimed at addressing the issues of insufficient generalization capability and insensitivity to subtle artifacts in existing methods. Firstly, a Multi-band Collaborative Forensic Filter (MCFF) is designed to cooperatively extract cross-difference, texture-statistical, and frequency-domain features. Secondly, a Dual-path Feature Interaction Module (DP-FIM) and a Dual-path Coordinate Attention Module (DP-CAM) are designed to strengthen feature fusion and spatial-aware representation. Thirdly, a lightweight High-frequency Perception Module (HFP) and an efficient Hierarchical Dual-path Attention (HDPA) mechanism are embedded in the deep layers to precisely capture imperceptible generation traces. Experimental results on four public datasets (ForenSynths, GANGen-Detection, DiffusionForensics, and UniversalFakeDetect) demonstrate that the proposed method outperforms existing state-of-the-art algorithms in terms of generalization performance.