Active thermography-based detection of thermistor joints using difference shadow method and deep learning algorithm

The quality of solder joints between aerospace-grade thermistors and wires is crucial for on-orbit spacecraft operation and state prediction. Active infrared thermography is widely employed for the inspection of solder joint defects. However, weak defect features in infrared images limit the detection accuracy and efficiency. In this paper, the difference shadow method was applied to enhance defect features in infrared images, and an improved NeckPro-YOLO was developed for identification of the solder defects. In NeckPro-YOLO, the C2f structure was optimized to C2f-FasterBlock-EMA(CFE) to strengthen the extraction of key features. Then Concat was replaced by SDI (Semantics and Detail Infusion) to enhance feature fusion and reinforce consistent defect responses. Meanwhile, the efficient modules of Concat and C2f-AdditiveBlock-CGLU (CAC) were employed for multi-level feature aggregation and enhanced feature representation, respectively. The results showed that the difference shadow method improves the Precision by 6.1% compared to the original dataset. Moreover, the proposed NeckPro-YOLO improves Precision, Recall, mAP50, and mAP50–95 by 2.5%, 3.0%, 3.1%, and 1.1%, respectively, while reducing the number of parameters and GFLOPs by 10.4% and 19.2%, respectively. Consequently, NeckPro-YOLO provides a lightweight and accurate solution for infrared solder joint defect detection.