INDT-YOLOV11 detection model based on variable space pooling denoising for rapid identification of bubbles and delamination

Abstract Internal defects such as bubbles and delamination in 3D-printed carbon fiber composite components severely compromise the mechanical properties of these components. Infrared thermography is an effective technique for detecting internal defects in materials. However, traditional detection methods depend on manual inspection, which suffers from several limitations, including low efficiency, relatively high false detection rates, and slow processing speed. These shortcomings hinder their ability to meet the demands of industrial applications. This study proposes an model of Infrared Non-destructive Testing—You Only Look Once V11 (INDT-YOLOV11) rapid object detection model. Based on YOLOV11n, we present a VSP-C3k2 module incorporating variant spatial pooling to mitigate noise interference during feature fusion. The SaE mechanism from SENetV2 is integrated with the C2PSA module to form a SaE-C2PSA attention mechanism, enhancing defect feature representation. To improve detection speed, the original backbone network in YOLOV11n is replaced with a MobileNetV2 feature extraction network. A dataset containing bubbles and delamination defects was established using INDT based on halogen lamp to capture infrared thermal images. Experimental results demonstrate that compared with YOLOV11n, the proposed INDT-YOLOV11 model achieves 20.85% and 20.75% improvements in mAP50 and mAP50-90 metrics respectively, while reducing parameter count and computational load by approximately 38.59% and 39.06%. On GTX1050 hardware, INDT-YOLOV11 achieves 31.5 FPS with 23.05% speed enhancement. And INDT-YOLOV11 has achieved almost the same mAP50 as YOLOV8-worldv2 and YOLOX lightweight models with minimal parameter and computational complexity, making it highly competitive in models such as YOLO-SA, providing effective technical support for automated internal defect detection in 3D-printed carbon fiber composite components.