Enhanced Zero-Shot YOLOv10 for Multi-Class Tiny-Object Detection of Steel Surface Defects

Multi-class tiny-object detection technology has undergone decades of research, yet its performance and efficiency remain less than satisfactory. Traditional methods, which rely on high-resolution images, face inevitably increased computational costs as the size of images and features grows. Moreover, the limited availability and difficulty in acquiring small-object samples further complicate detection tasks. In this study, we propose an Enhanced Zero-Shot YOLOv10 (EZS-YOLOv10) method tailored for tiny-object detection, with a specific focus on detecting defects in steel materials. This approach incorporates adaptive multi-scale feature fusion to better handle small defects, alongside improved anchor mechanisms within dense connection and residual learning frameworks to enhance robustness and accuracy. Additionally, we introduce a Region Feature Synthesizer Network that leverages known class features to synthesize unknown class features, integrating feature alignment and cross-modal feature learning to achieve detection of unseen or few-sample defect types. To validate the efficacy of the proposed method, comprehensive experiments were conducted on the COCO and VisDrone datasets, which include a higher proportion of small objects, as well as a dedicated steel defect dataset. Results demonstrate that our method improves the detection mAP by 1.5 and the small-object mAP by 2.7 on the COCO dataset. On the VisDrone dataset, the mAP increased by 2.7, with significant improvements in the accurate recognition of rare samples. In the context of steel defect detection, the proposed method achieved superior performance in identifying various defect types, highlighting its practical applicability in industrial settings.