<p>Ceramic substrates are essential in semiconductor manufacturing, and precise defect detection is critical for maintaining product quality. However, defects in such substrates often exhibit low contrast, subtle inter-class variations, and high intra-class similarities, making detection challenging. Additionally, industrial datasets frequently suffer from annotation errors, such as incorrect or incomplete bounding boxes, further complicating detection accuracy. To address these challenges, we propose DAGG-YOLOv9 (Dual Attention Ghost Guided YOLOv9), an improved defect detection method grounded in the YOLOv9 framework. Firstly, we alleviate the effect of annotation errors by employing the Efficient Joint Intersection (EIoU Loss) technique, which improves the center-intersection (CIoU Loss) by considering shape, scale, and orientation. The model integrates a Ghost model module, combining a skip-join fusion operation and content-aware up-sampling (CARAFE), to boost performance while reducing parameters. Additionally, dual attention mechanisms A² and Efficient Channel Attention (ECA) are applied across multiple output headers to improve feature representation, particularly for poorly labeled data. The dual attention mechanism improves the DAGG-YOLOv9, which incorporates the Ghost module and other improvements, to enable precise identification of surface flaws on ceramic substrates. The experimental outcomes demonstrate that DAGG-YOLOv9 improves mean average precision (mAP) by 3.76%, reaching 58.05%, while enhancing detection accuracy for defects such as adhesions, scars, and cracks to 74.1%. Moreover, it reduces computational effort by 9.89%, demonstrating superior robustness and efficiency in the detection of ceramic substrate defects.</p>