A key points recognition and deformation correction method for a pointer instrument based on an improved RT-DETR model

Abstract Parameter identification accuracy will be greatly affected by the skew and rotating pointer instrument image captured in industrial environments. Regarding this issue, a deformation correction method for a pointer instrument based on an improved real-time detection transformer (RT-DETR) model is proposed. Firstly, the inner endpoints of the long scale line and the rotation center point are taken as the key points for ellipse fitting using the least-squares method. Then, based on the corresponding relationship between the rotation center and the fitted ellipse’s center, the elliptical and perspective transformation methods are used to correct the skew and rotating pointer instrument image in the first stage. Next, based on the symmetrical relationship between the zero and full scale marks of key points, the rotation transformation method is used to correct the skew and rotating pointer instrument image in the second stage. To improve the extraction accuracy and real-time performance of key points, the FasterNet convolutional gated module is constructed to optimize the backbone network of the RT-DETR model based on the Faster Net network. And the context-guided spatial feature reconstruction network-feature pyramid network, a spatial feature reconstruction network based on a context-guided network, is introduced to enhance the feature fusion of the RT-DETR model. Ablation and comparative experiments show that, compared with the original model, the optimized model, while maintaining the mAP@ 0.5–0.95 value unchanged, reduces the number of parameters by 25.72%, the computational cost by 34.21%, the model weight by 25.68%, and achieves a frames per second (FPS) value of 218.4 FPS. Instrument reading experiments show that the average relative reading error and the average absolute error using the proposed method are 2.033% and 0.427%, respectively, verifying the superiority, effectiveness, and generalizability of the proposed method.