<p>Casting defects in radiographic images often display unclear features, fuzzy contour boundaries, and uncertain locations, making accurate detection of defect targets highly challenging. Existing artificial intelligence (AI)-based defect-detection methods predominantly analyze single images, overlooking the <i>dynamic</i> observations encountered in practical applications. This limitation can lead to missed and false detections. To address these issues, this study proposes a low-semantic defect-detection method based on multi-image feature sequence fusion, combining the <i>dynamic</i> evaluation knowledge of technicians in analyzing radiographic testing images with the advanced YOLOv8 model. First, a multi-image decomposition strategy based on the gamma transform was developed to generate image sequences reflecting a pattern of dynamic changes, simulating the real-world process of defect evaluation. Second, a multi-image feature sequence fusion network was designed to capture spatiotemporal information, generating a feature matrix focused on the <i>dynamic</i> characteristics of defect targets, which was integrated with YOLOv8 for precise defect detection. Finally, degree-of-confidence similarity was employed to fine-tune candidate boxes, ensuring accurate localization of defect targets. Experimental comparisons with state-of-the-art methods, including RT-DETR, YOLOv10, and YOLOv11, demonstrate that the proposed method achieves superior defect-detection accuracy compared with existing methods, effectively addressing the issue of missed detections in casting defects. The proposed method achieves a recall rate of 96.71% and mean average precision of 77.09% for defect detection in casting-ray images, significantly enhancing the detection performance and localization accuracy for low-semantic defects.</p>