首页|Improving Residual Ceramic Core Detection via Adaptive Multi-scale Image Fusion Enhancement and Defect Data Augmentation in Neutron Radiography
Improving Residual Ceramic Core Detection via Adaptive Multi-scale Image Fusion Enhancement and Defect Data Augmentation in Neutron Radiography
Wang, Mr. Ning Zhang, Dr. Tian He, Dr. Linfeng Feng, Dr. Chengzhuo Wang, Miss Tianyun Wu, Dr. Meimei Zhao, Dr. Chenyi Qiao, Prof. Shuang
Improving Residual Ceramic Core Detection via Adaptive Multi-scale Image Fusion Enhancement and Defect Data Augmentation in Neutron Radiography
Improving Residual Ceramic Core Detection via Adaptive Multi-scale Image Fusion Enhancement and Defect Data Augmentation in Neutron Radiography
摘要
Neutron radiography (NR) has long been regarded as an important technique for detecting residual ceramic cores in aero-engine turbine blades because of its strong penetrating capability and high sensitivity to light elements. However, neutron radiographic images of turbine blades generally exhibit low contrast, and small residual ceramic core features are often indistinct and easily confused with background noise or structural textures. To improve the performance of automated defect detection, we developed an adaptive multi-scale image fusion enhancement algorithm and a defect data augmentation scheme for residual ceramic core detection in NR. First, building upon the tracer and doping approaches, a purely image-processing-based local contrast enhancement algorithm for residual ceramic core regions was further proposed to improve the visual perceptual quality of residual ceramic core images. Second, a residual ceramic core image data augumentation method based on stable diffusion inpainting was further proposed to address the scarcity of neutron radiographic images containing residual ceramic cores and the overwhelming proportion of images without residual ceramic cores using existing real residual ceramic core image data. Experimental results demonstrate that the proposed method effectively enhances the contrast of residual ceramic core images and improves the detection performance using YOLOv10 and RT-DETR as baseline models, achieving precision, recall, $\mathrm{mAP@0.5}$, and $\mathrm{mAP@0.5:0.95}$ of 91.44$\%$, 89.56$\%$, 87.60$\%$, and 63.73$\%$ for YOLOv10, and 89.26$\%$, 87.48$\%$, 88.13$\%$, and 65.57$\%$ for RT-DETR, respectively. These results demonstrate the potential of the proposed method for practical residual ceramic core detection in NR.