Abstract: Automatic fiber placement (AFP) technology is extensively applied in the manufacturing of complex components such as aircraft panels, fuselage sections, and engine inlets. Despite the high level of automation in the AFP process, there remains a significant gap in the availability of effective online defect detection methods. Traditional 2D image-based detection techniques often suffer from unclear defect features, leading to low accuracy, while 3D point cloud-based methods are costly in terms of both annotation and computation and often lack real-time processing capabilities. To address these challenges, this paper proposes a novel defect detection method leveraging 2.5D image information. A line laser sensor is rigidly attached to the AFP head to capture real-time contour information during the layup process. The collected contour data undergoes pre-processing steps, including filtering and baseline drift correction, before being transformed into a 2.5D image. Advanced deep learning-based object detection techniques are then employed, alongside a newly developed post-processing algorithm to accurately merge and filter defects within the images. Experimental validation conducted on an automated layup process for aircraft inlet ducts demonstrates the effectiveness of the proposed method in detecting various defects, such as gap, bridge, triangle, fold, and mix. The method achieves an overall false rate of 7.3% and miss rate of 4.1%. The detection efficiency reaches up to 143 frames per second (fps) with GPU acceleration and 13 fps on CPU, satisfying the real-time online detection requirements.