To address the challenges of small defect targets, multi-scale issues, and high reflectivity on the surface of the inner car door handle, we first tackle the problem of defect features being obscured during image acquisition due to surface curvature and mirror reflection by using a bowl-shaped light source and reducing the angle of the image acquisition surface. Then, recognizing the limitations of traditional RT-DETR models, such as poor detection accuracy and slow speed, we propose an improved RT-DETR object detection method. This method builds upon the RT-DETR framework, utilizing parallel dilated convolutions and the CA attention mechanism combined with convolutional re-parameterization in the backbone network to increase the receptive field and establish long-distance semantic information while improving the network inference speed. Additionally, extra detection layers are added to improve the network′s feature extraction capability for small object detection. In the multi-scale feature fusion stage, we use an improved BIFPN structure to enhance the model′s information interaction capability. Finally, ablation experiments show that, compared to traditional RT-DETR-based detection methods, our proposed improved RT-DETR method increases the mean Average Precision by 6.5%, achieves a detection speed 1.6 times that of the traditional model, and reduces the model′s parameter count to only 76.5% of the original network, validating the effectiveness of our proposed method.