A chip X-ray image bubble defect detection model combined with Dual-Former attention mechanism

• Published in: Measurement: Journal of the International Measurement Confederation
• Main Author: Li A.; Hamzah R.; Rahim S.K.N.A.
• Format: Article
• Language: English
• Published: Elsevier B.V. 2025
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85216873779&doi=10.1016%2fj.measurement.2025.116871&partnerID=40&md5=6675c1b67553df6fe12fb993d0a8b966

Bubble defects in chip packaging can have an impact on the stability and reliability of the chip. Existing defect detection methods exhibit limited performance in identifying small-sized bubble defects and are highly susceptible to low contrast and noise in chip X-ray images, leading to missed and false detections. To address these challenges, we propose YOLO-DFA, a defect detection model based on improved YOLOv8 framework, to improve the defect detection accuracy. First, a Dual-Former attention mechanism is introduced to improve local and global feature integration, addressing missed detections of small bubble defects and weakening meaningless noise information. Second, a C2-CS module replaces the C2f module in YOLOv8, reducing spatial feature redundancy and computational complexity. Third, an improved Neck network incorporates a 3D-CBS module into the PAFPN network, enhancing the recognition of low contrast targets by strengthening multi-scale feature fusion. DySample is used for upsampling to minimize feature detail loss. Experimental results on the CXray dataset demonstrate that the YOLO-DFA model surpasses YOLOv8 in Precision, Recall, mAP, and F1 Score indicators by 3.1%, 3.4%, 3.2%, and 3.2%, respectively, while achieving a detection speed of 145 FPS, meeting real-time detection requirements. On the ADP_MBT dataset, YOLO-DFA demonstrates strong performance in detecting other chip defects and exhibits notable generalization ability. © 2025 Elsevier Ltd