To address the impact of surface defects in aircraft skin—critical structural components—on overall structural performance

this paper proposes a deep learning detection network based on the RT-DETR model

aiming to enhance the accuracy and robustness of aircraft skin defect detection. In response to the challenges posed by defects of varying scales

morphologies

and complex distributions

a series of innovative techniques were adopted for optimization. First

in the feature extraction stage

multilevel convolution blocks (MCB) were introduced. Through multi-layer convolution operations

the discriminative power of features at different scales was strengthened

effectively capturing details at various levels. Second

in the feature fusion stage

a multiscale feature enhancement (MSFE) module was employed

using multi-size depthwise convolution kernels to build contextual information

thereby improving the network’s robustness and adaptability to multiscale defect features. Finally

in the regression stage

a shape-aware (Shape-IoU) optimization module was introduced

which optimizes the matching between bounding boxes and defect contours

significantly improving detection accuracy. Experimental results show that the proposed network achieved an mAP@0.5 of 94.8% on the Aircraft dataset

representing a 12.7% improvement over the original RT-DETR model. Furthermore

on the NEU-DET test set

the model attained an mAP@0.5 of 92.5%. These results validate the effectiveness of the model in improving both the precision and generalization capability of aircraft skin defect detection.